How does a text watermark work?
A first-principles investigation of how generation-time text watermarks work, from a weighted coin to Gemma, and where the evidence stops.

6 posts • 5 projects • 83 shared links
A first-principles investigation of how generation-time text watermarks work, from a weighted coin to Gemma, and where the evidence stops.

Four tools, a settings file, and full control - how I built my own AI coding agent with Pi instead of paying $200/month for a CLI that keeps changing.

Practical way of using colbert with ragatouille on modal labs

Practical strategies for implementing and optimizing Retrieval Augmented Generation (RAG) in LLM systems.



A config-driven pipeline for generating and filtering fine-tuning data with multi-source ingestion, Evol-Instruct, MinHash and LSH deduplication, and SQLite checkpoints.
Experiments with keyed text watermarking, statistical detection, model behavior, calibration, and the effect of editing on the signal.
A local search assistant that combines web search, document retrieval, agent tools, and language models with specialized ranking and chunking.
A Llama 2 7B model fine-tuned on 60,000 bilingual English-Gujarati pairs for low-resource language use cases.
A personalized chatbot for mental-health support and resources, built with Mistral AI and Whisper models during a hackathon.