Skip to content
View aamodbhatt's full-sized avatar
🎯
Focusing
🎯
Focusing
5 followers · 9 following

Achievements

Achievement: Quickdraw

Achievements

Achievement: Quickdraw

Block or report aamodbhatt

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
aamodbhatt/README.md

Aamod Bhatt, Machine Learning Engineer · Researcher, above stone pillars rising from a sea of cloud

Hi, I'm Aamod, a machine learning engineer. I build complete systems, from lie-detecting probes for LLMs to a dungeon you can play inside a diffusion model. Also a musician!

bhatt.aamod@gmail.com


Selected work

marionette. A polygraph for LLMs: linear probes that catch a model lying. 0.896 AUROC on held-out spontaneous lies. dreamcade: A playable 3D dungeon dreamed by a diffusion model, in the browser. 21.9 dB PSNR over 48-frame rollouts. quarry: A LangGraph text-to-SQL agent with an MCP server. 162 / 162 red-team cases passed. witness: A retrieval-augmented fact-checker for any webpage. 76.4% accuracy on 20k FEVER claims. bitgolf: Neural text compression with a tiny transformer. 1.50 bits per byte, 40% smaller than brotli. persuadables: Uplift modeling on a 64k-customer email A/B test. +7.4 pp visits for women’s-wear buyers. nandbox: An 8-bit CPU that plays Snake in the browser. 1,521 NAND gates, the only logic primitive. excelsior: A spreadsheet engine that runs a transformer in its formulas. 5.8e-15 max difference from PyTorch.

Research

OpenAI Parameter Golf: Test-time training with Muon on 8× H100. 1.1179 bits per byte, merged submission (3-seed mean).

Toolkit

Python PyTorch Hugging Face Transformers scikit-learn XGBoost LangGraph RAG MCP LoRA SQL DuckDB FastAPI Docker AWS TypeScript Go Rust

Earlier work
  • lack-of-invariance · A study of how LLM answers shift when a question is paraphrased
  • model-arena · A terminal arena where specialized agents debate in structured rounds
  • recursive-knowledge-engine · A RAG system that retrieves, answers, critiques and refines its own queries
  • ftuneprompt · LoRA experiments on when to fine-tune and when to prompt, with a CLI that recommends one

Popular repositories Loading

  1. XAU-USD-Forecaster XAU-USD-Forecaster Public

    Jupyter Notebook 1

  2. aamodbhatt aamodbhatt Public

    Python 1

  3. recursive-knowledge-engine recursive-knowledge-engine Public

    Python 1

  4. Mask-Detector-main Mask-Detector-main Public

    Jupyter Notebook

  5. opencv4nodejs opencv4nodejs Public

    Forked from justadudewhohacks/opencv4nodejs

    Nodejs bindings to OpenCV 3 and OpenCV 4

    C++

  6. cnlab cnlab Public

    Java