Skip to content
View tathadn's full-sized avatar
🎯
Focusing
🎯
Focusing

Block or report tathadn

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.

Maximum 250 characters. Please don't include any personal information such as legal names or email addresses. Markdown supported. This note will be visible to only you.
Report abuse

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

Report abuse
tathadn/README.md

Hi, I'm Tathagata Debnath (Dave)

PhD candidate in Computer Science at New Mexico State University. I build agentic AI systems and fine-tune foundation models on multi-GPU infrastructure. My research background is in bioinformatics and algorithm design (10 publications, 80+ citations, IEEE TPAMI, Nature, & IEEE/ACM TCBB).

Currently focused on autonomous code debugging agents, vision-language model adaptation, and self-improving LLM systems using MCTS, DPO, and LoRA/QLoRA on distributed H100 setups.


Featured Projects

CodeQ — Autonomous code debugging agent using Monte Carlo Tree Search + Direct Preference Optimization. The model explores fix strategies via MCTS, critiques them with dual-temperature self-evaluation, and improves each round through DPO fine-tuning. Built on Qwen2.5-Coder-7B across two H100 nodes (4-bit inference + bf16 training). Achieved 81.3% fix rate on DebugBench (up from 10% pre-refactor), with DPO pushing MCTS mode to 84%.

Parallel Multi-Agent Codegen — DAG-based multi-agent code generation system using LangGraph + native Anthropic SDK. Orchestrates parallel async workers that plan, generate, and review code through structured handoffs.

Self-Evolving Codegen — Extension of the multi-agent pipeline with an LLM-as-Judge evaluator and autonomous prompt evolution loop. The system iteratively rewrites its own generation prompts based on evaluation feedback.


Publications & Packages

  • 10 peer-reviewed publications | 80+ citations — Google Scholar
  • OptCirClust — Fast optimal circular clustering algorithm (CRAN R package)
  • CircularSilhouette — Circular silhouette index for cluster validation (CRAN R package)

Links

Website · Google Scholar · LinkedIn

Pinned Loading

  1. multi-agent-codegen multi-agent-codegen Public

    A multi-agent AI system that transforms natural language task descriptions into working, tested code using specialized LLM agents that plan, write, review, and test code collaboratively. Includes L…

    Python

  2. parallel-multi-agent-codegen parallel-multi-agent-codegen Public

    A multi-agent AI system where an orchestrator agent decomposes coding tasks into a dependency graph (DAG), dispatches parallel coder workers for concurrent code generation, then merges, reviews, an…

    Python

  3. self-evolving-codegen self-evolving-codegen Public

    A multi-agent code generation pipeline with a self-evolving tester that autonomously improves its test strategy over generations

    Python

  4. codeq codeq Public

    An AI agent that teaches itself to fix bugs — MCTS explores debugging strategies, DPO trains on what works. Pipelined across two H100 nodes: one for 4-bit inference and trajectory collection, one f…

    Python