I ended up using it for more general purpose things because being able to have a hands-free phone call with an AI turned out to be pretty useful.

It's offline now, but here's the code with all the stack and deployment info: https://github.com/kevingduck/ChatGPT-phone/

Edit: forgot to mention this was all running off a $35 raspberry pi.

But it sounds damn creative as a project.

And you're right, I'm not a sales guy. This project is for people like me who want a risk-free place to learn the basics of sales so that when I do talk to an actual human, I won't panic and freeze up like I always do.

Most high-level sales people rely on role play partners but that requires a pretty a big commitment. This would make a great product, imo.

Also (tip): Study, memorize and internalize a sales script for your product/service...along with the objection handlers and closing questions. Practice every single day. You'll gain massive confidence because you know exactly what you are going to say, every time.

That's turning out to be a valuable feature of LLMs in many areas. You can practice complex interactions with them without worrying about boring or annoying them. Even the most patient human teacher gets tired eventually. LLMs don't.

Please consider putting it in online.

For me there's none is more boring than marking/evaluating/grading manuscripts. I prefer hard labor like gardening or farming than doing that activities although I'm quite good at evaluating stuffs I think.

Can you please elaborate how you do this and based on what metrics/scheme/etc the answers are being evaluated?

It could work both for practice and for automating interviews as well

For the dictated recipes, I told him to dictate just "flat" the words and numbers. So that I had paragraphs of recipes.

For the scanned recipes, I used Google OCR (I found out it was the best one quality wise).

For both sets of recipes, I then used GPT4 to "format" the unformatted recipes into well formatted Markdown. It successfully fixed typos and bad OCR from Google.

We then pasted all that well formatted text into a big Google Docs, and added images. Using OpenAI image generation I generated images for each of the 250+ recipes. For some of them I had to manually curate it, given that some of the recipes are for typical Mexican food: For example there's a (delicious) recipe called "PibiPollo" that for the unitiated it may look like a stew, so I had to tell something like "large corn tamale with thick hard crust".

In the end, the book was pretty nice! We distributed digital copies within the family and everybody was amazed :) . I loved spending time doing that.

https://drive.google.com/file/d/1OGE-zfNHHDnALbhgmf3lykBjcSg...

It is in Spanish though.

Are you using the "Record Audio" action or something else? Ideally the shortcut would stop listening after a pause like the native Dictate feature does it. At a minimum Record Audio seems to require hitting spacebar to stop - not great but not terrible.

Wondering: How big is you monthly OpenAI bill when using all these tools? Only a few $$$, or is it higher?

The backend is a Python FastAPI that uses ChromaDB to store my resume and Q&A pairs, OpenAI, and Airtable to log requests and responses. The UI is Sveltekit.

I'm currently building a different tool and will apply some learnings to my Interactive Resume AI. Instead of Airtable, I am going to use LangSmith for observability.

I started writing and my Substack articles are also linked to via my website. I'm currently working on applying sentence window retrieval and that article will be out shortly. This is part of a #buildinpublic effort to help build my brand as well.

I've been unemployed since Sept as a Senior Software Engineer. The market is tough so I'm focusing on the above to help get employment or a contract.

I also built Resume Chatbot but using slightly different stack: Python, Langchain, Faiss as vector store, MongoDB to store chat logs and Streamlit for UI. Here is a link: https://www.artkreimer.com/resume/ or you can try it on streamlit https://art-career-bot.streamlit.app/. Code is available here https://github.com/kredar/data_analytics/tree/master/career_.... Great thread and I got some ideas for my next project. Thanks a lot everyone.

The actual questions I got did not provide a response that is to my liking. Most of that is due in part because I'm using gpt3.5 since gpt4-turbo is a lot more expensive, and I can learn a lot more by using an inferior LLM.

For example, using an llm router to analyze the query and route to a specific helper function with a specific prompt would be helpful. Sometimes a user starts with a greeting but the response is a pre-written "Sorry an answer cannot be found". Questions are typically grouped into a category such as skills, experience, project, personal (ie: where are you located), preferences (ie: favorite language), and general interview questions (ie: why should I hire you). Questions in categories can be better answered by using a different prompt and/or RAG technique.

I was contacted by a company recruiter for a small healthcare SaaS in California and had 3 interviews recently. When I looked up the job, only 7 people had applied in 2 weeks on LinkedIn. They are a very real company with very real people, but their job post is not getting seen (it's not a promoted post).

My next AI project will be to scrape LinkedIn jobs, analyze it for repost/promoted behavior, group it by consulting/headhunters vs company job post, eliminate duplicates, and filter based on my skillset and hard-no qualities (such as can't work if I live in California, must be in EST but I'm in PST timezone, requires Java experience, etc).

But still good point that there might be promoted jobs and non-promoted ones, maybe it's worth creating an own job scraper.

I'm currently also looking for a dev job. So you have 15 years of experience, live in California and struggle to find something? That sounds a bit demotivating to me lol, because I'm kinda half of all of that or a bit less.

I also like your LinkedIn analysis idea, should try that maybe, too.

* I built a real life Pokedex to recognize Pokemon [video] https://www.youtube.com/watch?v=wVcerPofkE0

* I used ChatGPT to filter nice comments and print them in my office [video] https://www.youtube.com/watch?v=AonMzGUN9gQ

* I built a general purpose chat assistant into an old intercom [video] https://www.youtube.com/watch?v=-zDdpeTdv84

Again, nothing terribly useful, but all fun.

1. sketch (in notebook, ai for pandas) https://github.com/approximatelabs/sketch

2. datadm (open source, "chat with data", with support for the open source LLMs (https://github.com/approximatelabs/datadm)

3. Our main product: julyp. https://julyp.com/ (currently under very active rebrand and cleanup) -- but a "chat with data" style app, with a lot of specialized features. I'm also streaming me using it (and sometimes building it) every weekday on twitch to solve misc data problems (https://www.twitch.tv/bluecoconut)

For your next question, about the stack and deploy: We're using all sorts of different stacks and tooling. We made our own tooling at one point (https://github.com/approximatelabs/lambdaprompt/), but have more recently switched to just using the raw requests ourselves and writing out the logic ourselves in the product. For our main product, the code just lives in our next app, and deploys on vercel.

The thing I'm working on now is AI mock interviewing. It's basically scratching my own itch, since I hate leetcode prep, and have found I can learn better through interaction. To paste a blurb from an earlier comment of mine:

I'm building https://comp.lol. It's AI powered mock coding interviews, FAANG style. Looking for alpha testers when I release, sign up if you wanna try it out or just wanna try some mock coding. If its slow to load, sorry, everything runs on free tiers right now.

I really dislike doing leetcode prep, and I can't intuitively understand the solutions by just reading them. I've found the best way for me to learn is to seriously try the problem (timed, interview like conditions), and be able to 'discuss' with the interviewer without just jumping to reading the solution. Been using and building this as an experiment to try prepping in a manner I like.

It's not a replacement for real mock interviews - I think those are still the best, but they're expensive and time consuming. I'm hoping to get 80% of the benefit in an easier package.

I just put a waitlist in case anyone wants to try it out and give me feedback when I get it out

Gonna apologize in advance about the copywriting. Was more messing around for my own amusement, will probably change later

Thanks for signing up, will send out an email once its ready to take for a spin!

Runs on a local LLM, because even using GPT3 costs would have added up quickly.

Currently requires CUDA and uses a 10.7B model but if anyone wants to try a smaller one and report results let me know on github and I can give some help.

https://github.com/thomasj02/AiFilter

Built entirely on Vercel & OpenAI. Took about a day, hardest part was configuring Sign In With Google. Had several dozen candidates use it, saved a lot of time and helped prioritize conversations.

I just did a brief writeup about it yesterday: https://www.linkedin.com/pulse/i-built-ai-hiringscreening-as...

How did they fare compared to candidates that went through the chat process?

I believe a large company like Meta, or any of the other companies with messaging platforms, would find this valuable. Especially because they will be fined by the UK for fraud that takes place on their messaging services.

- Site: https://emergingtrajectories.com/

- GitHub repo: https://github.com/wgryc/emerging-trajectories

I've helped a number of companies build various sorts of LLM-powered apps (chatbots mainly) and found it interesting but not incredibly inspiring. The above is my attempt to build something no one else is working on.

It's been a lot of fun. Not sure if it'll be a "thing" ever, but I enjoy it.

Ideally, I’d like the agents to then participate in prediction markets or “superforecasting” groups to use actual human predictions as baselines.

There's also Moss[3], a GPT that acts as a senior, inquisitive, and clever Go pair programmer. I use it almost daily to help me code and it has been an huge help productivity-wise.

[1] https://git.sr.ht/~jamesponddotco/allalt

[2] https://ollama.ai/

[3] https://git.sr.ht/~jamesponddotco/moss

Powered by whisper-timestamped [1] using a model trained by the local tech university TTÜ [2]

And it just… works! (with some tweaks and corrections)

[1] https://github.com/linto-ai/whisper-timestamped

[2] https://huggingface.co/TalTechNLP/whisper-large-et

[1] https://just-tell-me.deno.dev/

[2] https://github.com/franekmagiera/just-tell-me

Primarily it was a PoC to see if a document based chatbot could work without crossing trust boundaries by calling out to untrusted APIs. It only makes calls to localhost.

If you’re familiar with the novel you will be pleased to know that the chatbot ended a recent answer with, “I must go now as I have an appointment with my chamber pot and I wouldn’t want to keep it waiting.”

[1]https://github.com/FlowiseAI/Flowise

[2]https://lmstudio.ai/

[3]https://llama.meta.com/

[4]https://ollama.ai/

[5]https://www.gutenberg.org/ebooks/1079

Everything runs on a Mac Mini with the M2 Pro CPU/GPU and Mac OS Sonoma.

Stack is a combination of TypeScript (Next / Node) + Python with a pretty simple deployment setup right now (GHA -> Container -> Cloud Run).

The layout is clean, and it's fast. The summaries are solid as well.

We’re processing the top podcasts in many genres every day (currently thousands of daily episodes) and running them through our pipeline.

From this we’ve made a semantic search engine, for example: https://www.podengine.ai/podcasts/search?search_term=Should+...

We’re soon going to improve and summarise the responses from the raw embeddings in a few ways. Would love some feedback on the experience.

We have also opened up a keyword alerting feature to alert folks when they’ve been talked about in an episode.

Have you considered using embeddings to show similar podcasts?

I initially built it using llama.cpp for offline LLM inference, but soon discovered mlc-llm and moved to using it, because the latter is way faster and flexible.

[1]: https://apps.apple.com/us/app/private-llm/id6448106860

Ultimately I wanted a whole marketplace where anybody can create a tour and then sell it.

But the process of creating the tours was quite laborious.

So to speed this up I fed GPT-4 information about local points of information and had it write the questions and the multi choice answers. It also wrote some narrative bits as various personas. For example, there was a Christmas hunt where GPT4 played the part of an elf and came up with a theme about Santa needing to recruit you to be a new elf, once you’d answered all the various clues etc.

Front end is React Typescript, backend is Net Core Web API on Linux with MySQL under EF Core and also integrations with GPT4 and Stripe.

It’s hosted at treasuretours.org

Only superusers can access the AI tools right now because cost, but you can try out some of the pre-made hunts which were partially AI generated.

I have been hacking together a poor-man's crunchbase that's fueled by GPT.

React / Python / Supabase. The most interesting piece thus far has been the success of the self-correcting loops through GPT. At each turn basically feeding the results back to another 3.5 prompt that is only about reviewing quality. I found that with these loops you can get solid results without having to use the more expensive GPT4 API.

(Also loving all the projects in this thread)

I've built some abstract content development tools, generally focused on building larger content somewhat top-down (defining vibes, then details).

I'm working on a general project helper using the GPT-Vision, voice, and regular GPT. You setup the camera above your workspace, work on paper, and chat with the LLM while you do it. I think there's a lot of potential, but the voice stuff is quite hard to deal with... there's just a ton of stuff happening in parallel, and I find it very hard to code something reliable.

The stack I use is all in the browser, generally Next.js, Preact Signals, and my own code to call into GPT, Whisper, etc. I like having everything available for inspection, and I generally keep all the working bits visible somewhere. (This can be overwhelming when other people see it.)

But I haven't gotten over the deployment hump... the cost and complexity is a challenge. I've used Openrouter.ai recently in a project, and I think if I leaned on that more completely I'd find the release process easier.

Here's the story:

At first I was building a tool for stock analysis- the user writes in free language what companies they want to compare, along with a time period, and their requested stocks show up on a graph. They can then further reiterate on it- add companies, and change range all in free language (I had many more analysis functions planned). Following some unique dev challenges I've found- I ended up not releasing the product (possibly will sometime in the future..), and switched to work on a dev platform to help with these challenges.

I was using what I called 'LLM structured task'- basically instructing the LLM to perform some task on the user input, and outputting a json that my backend can work with (in the described case- finding mentioned companies and optional time range, and returning stock symbols, and string formatted dates). The prompting has turned out to be not trivial, and kind of fragile- things broke with even minor iterations on the prompt or model configurations. So- I developed a platform to help with that- testing (templated) prompt versions, as well on model configurations on whole collections of inputs at once- making sure nothing breaks in the development process (or after). * If you're interested, welcome to check it out on https://www.promptotype.io

AI concierge for my parents’ vacation rental. Mostly just pulling info from the guest binder, but I’ve also started using some local guides to give better suggestions. Built with NextJs and deployed on Vercel (was really easy and they have a generous free tier).

It's macro + calorie tracking over text message. You just text what you eat and it matches against a food database to estimate your food intake. It's basically an easier alternative to MyFitnessPal.

My stack is OpenAI on Azure, Vercel, Convoy, FatSecret API, Postmark, NextJS.

https://peekly.ai

It’s basically an LLM-based RAG that works over the best blogs and websites covering any topic you provided during onboarding.

This is in firefox with and without UBlock Origin.

Errors: https://i.imgur.com/N28wnVY.png

{"data":{"error":"Imgur is temporarily over capacity. Please try again later."},"success":false,"status":403}

The tool still needs a trust mechanism and a coherent incremental publishing strategy to be able to operate in a public fashion. Right now, running one node using my RTX 3060 it would take 1.2 years to do one split of the C4 dataset.

https://arxiv.org/abs/2401.16380

https://www.emergentmind.com/papers/2401.16380

https://github.com/gardner/gsd

Deployment is usually FastAPI for business logic, Langchain or MS/Guidance library, LLM hosted via. HF-TGI server

Then things like:

“Fix My Japanese” - uses LLM to correct Japanese grammar (built with Elixir LiveView): https://fixmyjapanese.com

It has different “Senseis” that are effectively different LLMs, each with slightly different style. One is Claude, one is ChatGPT.

Or a slack bot that summarizes long threads:

https://github.com/dvcrn/slack-thread-summarizer

Followed you on Github, I'm looking at moving (back) to Japan in two years, likely to start a bootstrapped startup business and always good to have dev friends!

Sadly, every time I tried the fix my japanese page it just says the text I inputted wasn't Japanese. Maybe next time it'll work for me.

What kinds of prompts are using for the file renaming scripts?

Basically, I want to write a book without having to type out the whole thing. I got the dictation idea from an episode of Columbo.

It is very much a work in progress and a proof of concept for another writing tool I want to make.

https://orderly.cmgriffing.com/

https://github.com/cmgriffing/orderly

My mac mini is heavily constrained due to only having 2 proper cores and not much RAM. So the smallest models run best. The quantized tiny runs better than the regular tiny simply due to memory pressure.

So, in my testing on my mac mini it tends to take about 30% more time to process than the audio clip that was recorded. I added a specific warning that lets the user know to reduce their thread count if the processing time takes longer than a specific threshold.

Some of my stream viewers report it being pretty fast though. (much faster than what they see on my mac mini).

But I found that LLMs are often wrong and hallucinates, so I have to double check with google or other resources.

So I built a google and chatgpt alternative to answer any question and hallucinations are more obvious. I do this by using by multiple LLM's including search enabled ones i.e. GPT4, Gemini, Claude, Perplexity, Mistral, and Llama.

It's been growing healthily https://labophase.com

I made it available to the public aisearch.vip

https://github.com/codevideo/codevideo-ai

My goal is to definitely NOT generate the course content itself, but just take the effort out of recording and editing these courses: you provide (or get help generating) the stuff to speak and the code to write and the video is deterministically generated) The eventual vision is to convert book or article style text to generate the steps to generate the video in an as-close-as-possible-to-one-shot.

I also leverage Eleven Lab's voice cloning (technically not an LLM, but impressive ML models nonetheless)

For anyone more curious, I'm wondering if what I'm trying to do is in general a closed problem - to be able to generate step by step instructions to write functional code (including modifications, refactoring, or whatever you might do in an actual software course) or if this truly is something that can't be automated... any resources on the characteristics of coding itself would be awesome! What I'm trying to say is, at the end of the day code in an editor is a state machine - certain characters in a certain order produce certain results. Would love if anyone had more information about the meta of programming itself - abstract syntax trees and work there comes to mind, but I'm not even sure of the question I'm asking yet or trying to clarify at this point.

Unfortunately only for German, but I plan on expanding the languages soon.

https://vokabeln.io

Tech stack: - The app is in Flutter. - Backend I'm nodejs TS. - GPT4 for generation of sentences and explanations - GCP text-to-speech for audio

It doesn't do a ton, but it's kinda cool. Feel free to fix/add anything https://github.com/k-zehnder/gophersignal

GPT-4 excels as a translator, but it often encounters issues with content warnings and formatting errors when translating entire subtitle files via ChatGPT. The solution is straightforward: divide the subtitle file into sections, focusing solely on translating the text and disregarding the timestamps. While it's feasible to have ChatGPT maintain the correct format, I've observed a decline in translation quality when attempting this in a single pass. My preferred approach is a two-phase method: first, translate the text, and then, if necessary, request ChatGPT to adjust the formatting.

The webapp splits the srt file into batches of 20 phrases and translates each batch. It also allows for manual correction of the final translation.

Ah and it's also serverless: you input your OpenAI token & select the model of your choice and the webapp makes the requests to OpenAI directly.

- https://github.com/iloveitaly/sql-ai-prompt-generator generate a ChatGPT prompt with example data for a sqlite or postgres DB

- https://github.com/iloveitaly/conventional-notes-summarizati... summarize notes (originally for summarizing raw user interview notes)

- https://mikebian.co/using-chatgpt-to-convert-labcorp-pdfs-in... convert labcorp documents into a google sheet

- https://github.com/iloveitaly/openbook scrape VC websites with AI

It's just a hodgepodge of prototype scripts, but one that I actually used on a few occasions already. Most of the work is manual, but does seem easily run as "fire and forget" with maybe some ways to correct afterwards.

First, I'm using the pyannote for speech recognition: it converts audio to text, while being able to discern speakers: SPEAKER_01, _02, etc. The diarization provides nice timestamps, with resolution down to parts of words, which I later use in the minimal UI to quickly skip around, when a text is selected.

Next, I'm running a LLM prompt to identify speakers; so if SPEAKER_02 said to SPEAKER_05 "Hey Greg", it will identify SPEAKER_05 = Greg. I think it was my first time using the mistral 7b and I went "wow" out loud, once it got correct.

After that, I fill in the holes manually in speaker names and move on to grouping a bunch of text - in order to summarize. That doesn't seem interesting at a glance, but removing the filler words, which there are a ton of in any presentation or meeting, is a huge help. I do it chunk by chunk. I'm leaning here for the best LLM available and often pick the dolphin finetune of mixtral.

Last, I summarize those summarizations and slap that on the front of the google doc.

I also insert some relevant screenshots in between chunks (might go with some ffmpeg automatic scene change detection in the future).

aaand that's it. A doc, that is searchable easily. So, previously I had a bunch of 30 min. to 90 min. meeting recordings and any attempt at searching required a linear scan of files. Now, with a lot of additional prompt messaging I was able to:

- create meeting notes, with especially worthwile "what did I promise to send later" points

- this is huge: TALK with the transcript. I paste the whole transcript into the mistral 7b with 32k context and simply ask questions and follow-ups. No more watching or skimming an hour long video, just ask the transcript, if there was another round of lay-offs or if parking spaces rules changed.

- draw a mermaid sequence diagram, of a request flowing across services. It wasn't perfect, but it got me super excited about future possibilities to create or update service documentation based on ad-hoc meetings.

I guess everybody is actually trying to build the same, seems like a no-brainer based on current tool's capabilities.

So I view whisper as sota with excellent accuracy.

Now, for the type of transcription I need speaker discerning is much more valuable than accurate to the point translation: so it will be summarized anyway and that tends to gloss over some of errors anyway.

That said, pyannote has also caught me off guard: it correctly annotated lazily spoken "DP8" with non native speaker accent.

It looks really good

The stack is mostly python running locally, and calling the OpenAI API (although we have plans to support offline models).

For better visual understanding, we use a custom fork of Set-of-Mark prompting (https://github.com/microsoft/SoM) deployed to EC2 (see https://github.com/OpenAdaptAI/SoM/pull/3).

https://candle.fi/gpt

Our backend stack: - AWS - SST - TypeScript

Our clients:

- Next (web) - Vanilla React Native (mobile)

OpenAI's App Store announcement is what got us interested in building w/ LLMs.

We have an /updates page & blog, as well as links to GitHub. I figured finding out more about myself and my co-founder was pretty easy.

1. Analyze calories/macronutrients from a text description or photo

2. Provide onboarding/feedback/conversations like you'd get from a nutritionist

https://www.fatgpt.ai/

My stack is Ruby on Rails, PostgreSQL, OpenAI APIs. I chose Rails because I'm very fast in it, but I've found the combination of Rails+Sidekiq+ActionCable is really nice for building conversational experiences on the web. If I stick with this, I'll probably need a native iOS app though.

Vendor stack is: GitHub, Heroku (compute), Neon (DB), Loops.so (email), PostHog (analytics), Honeybadger (errors), and Linear.

Step 1: Is it a hot dog or not hot dog? https://www.youtube.com/watch?v=ACmydtFDTGs

I'm glad someone is keeping the dream alive!

- In my chipotle bowl, it can tell if I had brown rice vs white rice

- In my In-n-out, it can tell if I got it protein style

It struggles with accurate weights/volumes but I'm excited about where this is going.

The more notable one was experimenting with LLMs as high level task planners for robots (https://hlfshell.ai/posts/llm-task-planner/).

The other is a golang based AI assistant, like everyone else is building. Worked over text, had some neat memory features. This was more of a "first pass" learning about LLM applications. (https://github.com/hlfshell/coppermind).

I plan to revisit LLMs as context enriched planners for robot task planning soon.

And I'm working on a webapp that is a kanban board where LLM and human collaborate to build features in code. I just got a cool thing working there: like everyone, having LLM generate new code is easy but modifying code is hard. So my attempt at working on modifying code with LLM is starting with HTML and having GPT-4 write beautfulsoup code that then makes the desired modification to the HTML file. Will do with js, python via ast, etc. No link for this one yet :) still in development.

So, far, none of the 7 tested models were able to win even one of the easiest text adventures. I tried many prompting techniques. But only GPT-4 was able to play through the first half of the game.

[0] https://github.com/s-macke/AdventureAI

But Zork wouldn't be a very accurate measure of skill because GPT definitely knows Zork. Unfortunately the emulator (https://github.com/DLehenbauer/jszm) doesn't work with most games newer than Zork. I haven't revisited the code with newer GPT models either.

However, my code run a newer version of the Z-machine. So Zork and many other text adventures will work. I have not tried many other games though.

I did find giving GPT some hints about the known commands helped a lot, and I put in some detection of error messages and kept a running log of commands that wouldn't work. Getting it to navigate the parser is kind of half of the skill of playing one of these games. It would be interesting to have it play some, then step back and have it reflect and enumerate things about how the play itself works.

Of course you try even more sophisticated techniques than mine. I tried the ReAct pattern and virtual discussions. So far, he always stumbles at the same place in a critical understanding of the text. And I tried exactly this critical step dozens of times.

You will understand the issue yourself, once you play the game yourself. It just takes 20 minutes and is very easy:

https://adamcadre.ac/if/905.html

A long time ago I did some exercises in "classical planning algorithms", which all feel very like the early part of this game. I.e., how do you get ready to leave if you have to shower, and can't do that with clothes on, etc. A similar planning example involved changing a tire (opening the trunk, removing lug nuts, etc). It was surprisingly difficult to make an algorithm that could figure it out! You could search the state space given the transitions, but it exploded with what was effectively lots of dead ends; obvious to me as a human, but not to the algorithm. Which is to say that this is a harder problem than it might seem.

I chose this game, because the game just helps you, at the every step, what you have to do next. Not much to try out. Just the narrative changes. One time, you have to go to work and one time you have to flee.

Other text adventures are even more problematic. I saw GPT-4 trying for dozens of steps in the "The Hitchhiker's Guide to the Galaxy" adventure just to turn on the lights. And this just the first command you have to get right in the game.

Also, hello HN! If you are interested, use this promo code for 50% off your first purchase ;)

  HELLOHACKERNEWS

****

Project 2 - I also built a YouTube summarizer for individual video called Summary Cat (https://www.summarycat.com). It is not open source for now. The stack is very similar to project 1.

****

And yes I like summarizing YouTube videos:)

It is a Next.js application, calling OpenAI’s API using a plain API route.

[1] https://spliit.app

[2] https://spliit.app/blog/announcing-receipt-scanning-using-ai

A Decentralised AI App store with cross border micro transactions.

You will be able to sell your LLM output (could be multi modal) for dollars or you decide. (LLMs working on your infra, you can keep weights for yourself forever.)

https://dev.invoker.network/share/9/0 (Dev environment is ready).

https://dev.invoker.network/share/9/1

1. Games you can play with word2vec or related models (could be drop in replaced with sentence transformer). It's crazy that this is 5 years old now: https://github.com/Hellisotherpeople/Language-games

2. "Constrained Text Generation Studio" - A research project I wrote when I was trying to solve LLM's inability to follow syntactic, phonetic, or semantic constraints: https://github.com/Hellisotherpeople/Constrained-Text-Genera...

3. DebateKG - A bunch of "Semantic Knowledge Graphs" built on my pet debate evidence dataset (LLM backed embeddings indexes synchronized with a graphDB and a sqlDB via txtai). Can create compelling policy debate cases https://github.com/Hellisotherpeople/DebateKG

4. My failed attempt at a good extractive summarizer. My life work is dedicated to one day solving the problems I tried to fix with this project: https://github.com/Hellisotherpeople/CX_DB8

2) https://amiki.app - practise speaking French, Spanish, German or Italian with a 3D partner. Flutter web with Whisper and my own rendering package.

https://github.com/StuartRiffle/ragtag-tiger

https://github.com/psugihara/FreeChat

I'm also working on a little text adventure game that I hope to release soon.

Now with LLMs it’s simple to extract structured data from emails.

I built [Orderling](https://orderl.ing) that is basically a CRM for your orders. It uses OpenAI api to extract the order information and automatically adds it.

The results were actually hilarious... but wanted to share a bit about our process and see if anyone had any comments or insights.

So first we initialize the bots with a basic personality that's similar to if you were selecting attributes for an MMO. Things like intelligence, toxicity, charisma and the like. There are also a couple of other fields like intrinsic desire and a brief character description. These are fed to the model as a system prompt with each inference.

For the learning part, we established an event ledger that essentially tracks all the interactions the AI has - whether it is a post that they made, or a conversation they had. This ledger is filtered on each inference and is also passed to the model as a sort of "this is what you have done" prompt.

Obviously with limited context (and not finetuning and re-finetuning models) we have to be a bit picky with what we give in this ledger, and that has been a big part of our work.

Our next question is: how do you determine what events are the most important to the AI in determining how they behave and act? It's been interesting!

The platform is anotherlife.ai for those curious!

https://github.com/pramatias/documentdf

Something i want to test, is how much documentation is needed, for the machine to infer the rest of it. Something like, one sentence of human documentation + code, how much can LLM infer and describe the code as accurately as possible. Does it need two sentences? 3? We'll see.

You just export and upload a WhatsApp conversation and it will learn the personality AND voice of your conversation partner. You can send/receive text or voice messages; It was pretty damn spooky to actually have a voice conversation back and forth with an AI standing in for my "friend"

As for the stack, I have Supabase and Typescript on the frontend, python on the backend and k3's as a cluster for my apps (can recommend this if you want to get devops-y on a budget). Next time, I'll just go pure Typescript since python really doesn't add much working this far away from the base models.

In fact, when we showcased our prototype of the "improve my writing" feature that we quickly put together in one week in early 2023 to a few select clients, their feedback was very enthusiastic. However, it took several months to bring it into production due to several bureaucratic hurdles: clearance from the legal department, the product owner delaying the release because of other priorities, and so on.

Now that the first feature has received a lukewarm reception because of the delayed release, we have neither a dedicated team nor a budget for adding more LLM-based features. Implementing proper and useful RAG is more complex and requires a certain level of expertise (vector DB integration, chunking strategy/indexing, tricks like HyDE, reranking, etc.), compared to just using the command "hey ChatGPT, improve this: %s." It is now unlikely that we'll have anything cool anytime soon without the support of the CEO or product owner (and they probably believe I proved it was indeed a fad). Most teams are currently busy with ordinary features from our backlog that do not require LLMs, and no one cares anymore.

We built it in Kotlin with Ktor server, htmx and tailwind. It uses a mixture of models, including gpt4-turbo, gpt4-vision and gemini-pro-vision. It's deployed using Kamal on bare metal.

Example canvas that provides a roundup of Apple Vision Pro reviews: https://jumprun.ai/share/canvas/01HNXB2K3GM7KPRP45Y2CVVJSC

Our learn more page with some screenshots to show creating a canvas: https://jumprun.ai/learn-more

It's a free closed beta at the moment to control costs, but let me know if you'd like an invite.

The last app, the only one that was deployed anywhere, is https://catchingkillers.com This app is a simple murder mystery game where the witnesses and the killer are ChatGPT bots. The first two stories are complete and active, the third is not complete yet. The first story of the working two is taken from another murder mystery group game https://www.whodunitmysteries.com/sour.html. The second story was highly influenced by ChatGPT.

It's a bit rough because I didn't spend too much time on it, but if anyone does signup to play, I'd love to hear feedback.

I would also in future try to make it generic so that it can crawl any website and store new contents in vector databases. Response to user query then can be returned by combining the vector search and llm

https://link.springer.com/chapter/10.1007/978-3-031-28238-6_...

Financial earnings calls are important events in investment managements: CEOs and CFOs present the results of the recent quarter, and a few invited analysts ask them questions at the end in a Q&A block.

Because this is very different prose from news, traditional summarization methods fail. So we pre-trained a transformer from scratch with a ton of high-quality (REUTERS only) finance news and then fine-tuned with a large (100k sentences) self-curated corpus of expert-created summaries.

We also implemented a range of other systems for comparison.

It looks through past transcripts, topics, view counts, and other metadata so users can quickly learn what a Youtuber is all about.

AgentX (https://theagentx.com), an LLM chat support app is one of the projects I built on this framework. It is a self-updating customer support agent that is trained on your support docs. Not only does this answer your customer questions, it provides summaries of the queries so you get a sense of where your product and/or documentation is deficient.

I’m working for an edTech company. Some students prefer video. So I built a Django app that takes a block of text and formats it into a set of slides, each with a title, some bullet points, an Dalle-3 generated image, and a voiceover.

It then compiles that all into a video.

The stack is simple, preact in the fronted with a custom framework on top and bun on the backend calling OpenAI, I may port it to rust in the future.

I plan to try local LLMs when I have some free time.

For now each users runs the application locally with their own keys[3].

[1] https://www.youtube.com/watch?v=nS1wsif3y94

[2] https://www.youtube.com/watch?v=f-txlMDLfng

[3] Alpha software, check the readme: https://gloodata.com/download/

Example 1: i get an email from a potential customer that says they want [product A]. I can forward that email (or call notes) to salesforce (or somewhere) and it will understand the preference and the relevant customer and update that customer's profile.

Example 2: In a B2B context, lets say my customer is a company, and there is a news article about them. I could forward a link to the article to the LLM and it would understand that the article is about a customer, and append that article and key info about it to my saleforce record for that customer. The news item becomes an object that is linked to that customer (for call context, better sales targeting, profiling, etc).

Can someone help me build that?

So far this is being used for:

- Sales -> guiding new recruits during more complex client calls

- HR -> Capturing respones during screening interviews

If you'd like to try this out feel free to DM me or email me at andrew at sightglass.ai, we're looking for more testers!

https://domsy.io

Currently running on my little digital ocean droplet. Stack is javascript/python.

https://domsy.io/share/ddf54149-5de9-4f3a-b936-f007a451c0b5

It is still a work in progress (early beta), but you can check it out at https://www.bonamiko.com

Currently I have mainly been using it as a tandem conversation partner for a language I'm learning, but it can be used for many more things. As it is right now, you can use it to bounce ideas of, practice interviews, and help answer quick general questions. You just need to tell it what you want.

The stack is a Next.js application hosted on Vercel using Supabase for the backend. (There is also some plumbing in AWS for email and DNS.) It is automatically deployed via GitHub actions.

To answer your question more specifically,

Pro Bonamiko:

  - Faster average first response latency (but higher first audio latency since OpenAI uses a ding). This is the main focus currently, reducing latency as much as I can. I'd like to be able to avoid the ding, but we'll see how low I can get it.
  - Can be used anywhere with a browser, OpenAI requires a mobile app installed. (I.E. Desktop support)
  - In the future we can support deeper customization since we are focused on the audio medium. As soon as you have to run a function in the ChatGPT app there is a long response latency, which could easily be fixed by something as simple as the AI saying "Let me perform a search to get the details"

  - Nice animation
  - Already has built in tool support such as web search
  - Supports language switching automatically between messages, Bonamiko requires manually changing the language

Since it is python library, we deploy it to pypi. But for using it on my own, I am using H100 linux server on the torch docker & CUDA. Running it needs only vim and bash. And plus, for running local model I love VLLM. I make my own VLLM Dockerfile and use it for deploying local model in 5 minutes.

FYI : Borrowing whole H100 instance is really expensive, but in my hometown, the government support us the instance for researching AI.

I think what will be really powerful is to have a registry for plugins and agents that can be easily installed in the system. Sort of like WordPress in that way. Also similar to an open source GPT store.

https://github.com/runvnc/agenthost

I believe the are several variations of this type of idea out there.

This project is a chatroom application that allows users to join different chat rooms, send messages, and interact with multiple language models in real-time. The backend is built with Flask and Flask-SocketIO for real-time web communication, while the frontend uses HTML, CSS, and JavaScript to provide an interactive user interface.

demo here supports communication with `vllm/openchat`:

* http://home.foxhop.net:5001

[0]https://www.zigpoll.com

I’m curious: was there one method that improved the accuracy/relevance of the answers the most?

Also, are you using Langchain, Llamaindex, or something else?

It should be cheap enough to deploy that it can be applied to relatively low-value content like video meeting recordings, so it can’t spend a lot of expensive GPU time analyzing video frames.

It also needs to be easily customizable for various content verticals and visual styling like branding and graphics overlays.

And everything is meant to be open sourced, so that’s fun!

I wrote about it on my employer’s blog here:

https://www.daily.co/blog/automatic-short-form-video-highlig...

The stack is: 1. TypeScript/Node/tRPC/Postgres/Redis/OpenAI on the backend 2. SolidJS/Crxjs/tRPC on the front end 3. Astro for the docs/marketing site

And deployment is currently through render.com for the databases and servers, and manually via a zip file to the Chrome webstore for the extension itself.

[0] https://smudge.ai

https://humanornot.so/

Heavily inspired by https://humanornot.ai/ (which was a limited time research by Ai21 Labs), now the project is on its own path to be more that just a test.

My work is to make AI chats sound like real humans and it's shocking how good sometimes the AIs are .

Even I as a creator, knowing everything (prompts, fine-tuning data, design, backend etc.), often can't tell if I'm speaking to human or designed by me AIs

A tool to RAG a github repo, so i can ask questions of how a certain library or tool works? Even better if it pulls in issues

Langchain and Llama Index both have classes to read a directory, if you git clone, and perform a RAG.

If you want it to scrape a github URL, there is a module for that too!

This starter tutorial will do RAG on a directory of files: https://docs.llamaindex.ai/en/latest/getting_started/starter...

Github Link: https://github.com/joiahq/joia

Benefits vs the original: - Easy to invite entire teams and centralize billing - Talks to any Large Language Model (eg: Llama 2, Mixtral, Gemini) - Collaborative workspace to easily share GPTs within the team, similar to how Notion pages work - Savings of 50%-70% vs ChatGPT's monthly subscription

Tech stack: NextJS, Trpc and Postgres. All wonderful technologies that have helped me develop at the speed of thought.

I've built this by using AI as the foundation for everything. I am using LLMs to classify information and extract structured data points for any webpage, or RAG for finding data.

Tech stack: - Mistral 8x7b and Perplexity API for data processing and GPT-4 input - GPT-4 for content output - pgvector in Supabase - LangChain for the pipeline and RAG stuff

It's a flutter app (in beta on Google play store currently) that uses OpenAI embeddings with Postgres pg_vector DB hosted in Supabase. Any poor matches go to Dalle3 for generation.

Our charity (I am vice-chair on the board) is hoping to use it as part of our program: https://learningo.org/app/

I consult to a law firm as their founder-in-residence. For fun, I trained Llama 2 on all the non-client data of the firm so that people could ask it questions like "Who are the lawyers in Montreal who litigate American securities laws, what are their email addresses and what time is it where they are?" It's a njs app running on linode.

It's extremely simple, but people seem to find it useful.

Check it out here: https://app.commonbase.ai/

It has been a huge help for me when working with certain open-source libraries.

We used Plasmo to build the chrome extension, React for the frontend, and currently OpenAI as the LLM provider.

Currently it only works with Gmail but we plan on adding other email providers as well.

Feel free to check it out: https://chromewebstore.google.com/detail/mysterian-ai-for-gm...

I also wrote PromptPrompt, which is a free and extremely light-weight prompt management system that hosts + serves prompts on CDNs for rapid retrieval (plus version history): https://promptprompt.io

  - Nextjs
  - Deno Deploy for hosting the apis 
  - Supabase - postgres / auth
  - Shadcn

It's really MVP, but I want to see if anyone is interested at all before I work on v2: creating gpts that come with databases!

  [1] https://textool.dev
  [2] https://create.t3.gg/

https://msty.app

The stack is react / cloud run / job queue / LLMs (several) / vector db.