PhD Candidate · Computer Science · NMSU

Tathagata
(Dave)
Debnath.

Building agentic AI systems and fine-tuning foundation models on multi-GPU infrastructure.

Also available: Bioinformatics CV · Resume

About

PhD candidate building AI systems that fix bugs.

I'm a PhD candidate at New Mexico State University building AI systems that autonomously find and fix software bugs. My current work combines Monte Carlo Tree Search, Direct Preference Optimization, and vision-language model fine-tuning — all running on distributed NVIDIA H100 GPU infrastructure.

My research background is in bioinformatics and algorithm design, with 10 peer-reviewed publications (80+ citations) in venues including IEEE TPAMI and IEEE/ACM TCBB, and two published CRAN R packages for clustering algorithms.

I'm targeting Summer 2026 AI/ML internships and full-time roles in agentic AI, LLM fine-tuning, and ML systems engineering.

Current Work

· CodeQ — MCTS + DPO debugging agent
· VisionTriage — multimodal bug triage
· Multi-agent code generation pipelines

Seeking

Summer 2026 AI/ML internships and full-time roles.

Get in touch →

Education

PhD in Computer Science

New Mexico State University · Expected 2026

Focus: Cyber Security, AI/ML, Agentic Systems

M.Tech in Computer Science

Tripura University · 2017

B.Tech in Computer Science

NIT Agartala · 2015

Featured Projects

AI systems and research software.

View All Projects

CodeQ

Autonomous Code Debugging Agent

Flagship

Self-improving code debugging agent using Monte Carlo Tree Search to explore fix strategies, dual-temperature self-critique to rank them, and Direct Preference Optimization to learn from its own exploration data. Built on Qwen2.5-Coder-7B across two H100 nodes.

81.3%

fix rate on DebugBench

84%

MCTS mode after DPO Round 2

MCTS DPO Qwen2.5-Coder-7B LoRA 4-bit Quantization HuggingFace TRL Docker Sandbox 2× H100

GitHub → Blog Post → Details →

Parallel Multi-Agent Codegen

DAG-Based Agent Orchestration for Code Synthesis

AI / LLM

Multi-agent code generation system using LangGraph with a DAG-based orchestrator that dispatches concurrent coder workers via asyncio. Extended version adds an LLM-as-Judge evaluator and autonomous prompt evolution loop where the system iteratively rewrites its own generation prompts based on evaluation feedback.

LangGraph Anthropic SDK asyncio LLM-as-Judge Prompt Evolution Docker LangSmith

GitHub (parallel) → GitHub (self-evolving) → Details →

VisionTriage

Multimodal Bug Report Triage

In Progress

Fine-tunes Qwen2.5-VL-7B-Instruct with QLoRA to triage software bug reports from screenshots + text descriptions. Text-only baseline benchmarked against SevPredict on the standard Eclipse/Mozilla dataset. Multimodal extension shows that adding screenshot context improves severity prediction accuracy.

Qwen2.5-VL-7B QLoRA Eclipse/Mozilla Dataset Rico Synthetic Data Gradio HuggingFace

Details →

Technical Skills

Tools and systems I work with.

AI/ML & Foundation Models

PyTorch, HuggingFace (Transformers, TRL, PEFT), LoRA/QLoRA, DPO, MCTS, Flash Attention 2, bitsandbytes (4-bit/8-bit), vLLM, Qwen2.5, DeepSeek, LangGraph, LangChain, RAG

Infrastructure & Systems

NVIDIA H100 (multi-GPU), CUDA, NCCL, FSDP, DeepSpeed ZeRO, Docker, Git, Linux, screen/tmux, W&B, SSH

Languages

Python, R, C/C++, SQL, Bash, JavaScript, LaTeX

Bioinformatics (Research)

Samtools, STAR, DESeq2, DEXSeq, GATK, rMATS, RSEM, Gviz, gprofiler2, MUMmer4, BUSCO, Singularity

Tathagata(Dave)Debnath.