Avery Carter 🇺🇸

Claude now connects to the tools creative professionals already use. With the new Blender connector, you can debug a scene, build new tools, or batch-apply changes across every object, directly from Claude.

It’s extremely good that Anthropic has not backed down, and it’s siginficant that OpenAI has taken a similar stance. In the future, there will be much more challenging situations of this nature, and it will be critical for the relevant leaders to rise up to the occasion, for fierce competitors to put their differences aside. Good to see that happen today.

Vibe coded a game character selection screen Everything here was made with AI tools Nano Banana: character design + UI Tencent Hunyuan3D: image to 3D Gemini Pro: UI More details ↓

@aidenybai This is really nice! Super fast

introducing repogrep.​com ultra fast codebase search for any public github repo where i found the React hooks source code in <10 sec

@xuelinda7 Looks cool, would love to test

Wanted to be able to semantically search for papers to read So I built it out :) (lmk if u wanna beta test)

@andreasklinger @bilawalsidhu @gravicle @janusch_patas this looks like @rubenfro

This might be the coolest use of Gaussian Splats in a movie i have seen in a while. Wait for 0:20 /cc @bilawalsidhu @gravicle @janusch_patas

A video on how I used @cursor_ai @langchain @OpenAI to build a tool that helps me work on my car.

@jordihays put a porsche(s) in it

Lease signed on new 4,000 sqf studio 🏁

If you happen to have an e46 m3 and love to wrench. Feel free to add your email here and I'll send out a blast when it's ready to test docs.google.com/forms/d/e/1FAI…

It's called Traccshun, and I have it up and running (locally lol). I'm working on the prompt engineering to improve responses and work out some edge cases. I have a small waitlist from the bmw community that's eager to try it out!

So as it currently stands, this tool is like a fancy AI-enabled ctrl-f for this huge pdf. Doesn't sound like much but it solves a real problem and I'm already seeing a ton of other directions I can go with this 👨🏽‍🔧

It fetches all document embeddings from the LangChain-managed ChromaDB vector store, reduces their dimensionality to 3D, and normalizes them for display. I used this tool for that: github.com/vasturiano/3d-…

For extra cool points I built a 3D visualizer that shows embeddings as nodes connected based on similarity. The embedding highlights in red after every query that finds a valid response.

For the chunking method I decided to do compent-level chunking instead of token-based. So for each page in my torque data, I loop through all components (e.g., bolts, mounts, etc.) and create a distinct document for each one. I figured this would help with retrieval

The PDF was too unorganized and dense to work with vanilla PDF text extraction. It had tables, diagrams, part numbers, and random annotations all through it. I came to learn this is problematic for any PDF with tons of information like this

Hell yea. Now this RAG stuff became much easier. The responses are more consistent. The assistant was able to find most of what I was asking for at this point. I'm feeling good. "Alexa play Olympian by Playboi Carti"

So I then used Gemini Pro's huge context window and gave it the 100 page PDF and the json object structure I wanted, and it spat out 4100 lines of beautifully structured data

I understood the PDF structure well since I've spent like 2000+ hours wrenching on my car. And I figured maybe the best thing to do here is structure the document into a JSON object that then could be chunked into embeddings and then referenced back to the PDF

Langchain actually notes in their docs: "answering over PDFs with complex layouts, diagrams, or scans -- it may be advantageous to skip the PDF parsing, instead casting a PDF page to an image and passing it to a model directly" I ended up going a different way though

After tinkering with @langchain for a few hours, I had a half working RAG example, but it really struggled with finding accurate answers. This lead me down a rabbit hole of optimizing unstructured data for RAG. My issue was how this PDF was structured. It was not working.

I used @cursor_ai to vibe code a AI-Enabled RAG assistant that helps me find torque specs on my beloved BMW E46 M3. I'm calling it Traccshun 🧵 🛠️ Stack: - Next.js - @langchain (RAG) - @OpenAI (embeddings + LLM) - ChromaDB (vector search)

Retrieval Augmented Generation (RAG) seemed like the best solution to making an AI assistant that referenced this document. So that's where I started off. I wanted to minimize hallucinations and only provide answers specifically referenced from the PDF.