role : ML/AI builder, researcher, and systems tinkerer
focus : agentic AI, LLM evaluation, model alignment, applied ML products
currently : building testable LLM workflows, AI assistants, and research-grade ML pipelines
style : ship useful systems, document the reasoning, keep experiments reproducible
I like projects where the model is only one part of the system: retrieval, data quality, evaluation, user workflow, infrastructure, and failure handling all matter. My work moves across agentic systems, language-model fine-tuning, physics-informed deep learning, reinforcement learning, and product-oriented AI.
| Track | What I Build | Strongest Evidence |
|---|---|---|
| Agentic AI | structured LLM workflows, ReAct-style agents, evaluator loops, MCP demos | mellea-IBM |
| Applied AI Products | end-to-end AI apps with backend, frontend, storage, and deployment | ReVival, AtomMail |
| Research ML | PyTorch experiments, scientific ML, model evaluation, public artifacts | Deeplense_GSoC |
| LLM Training | QLoRA, LoRA, DPO, summarization, preference optimization | DPO Alignment, Mistral Summarization |
| Foundations | language modeling, RL agents, implementation-first learning | CS336 Study Repo, DQN Space Invaders |
ReVival is an end-to-end circular-commerce system for routing returned and seller-listed products back into the market instead of landfill. The system uses a 7-agent pipeline for product grading, resale routing, pricing, buyer matching, trust passport generation, sustainability credits, and listing-error prevention.
What I owned
- Supply-side AI and AWS backend.
- FastAPI pipeline for product returns and community listings.
- AWS Bedrock vision-based condition grading.
- Photo/video defect detection, confidence buckets, mismatch detection, and grading guardrails.
- DynamoDB and S3 data layer with
boto3. - Titan multimodal embedding cache to reduce repeated AI calls on visually similar products.
Stack: FastAPI, Python, AWS Bedrock, DynamoDB, S3, boto3, Next.js, TypeScript
DeepLense contains my ML4Sci/GSoC evaluation work across gravitational-lens classification, observational lens finding, and physics-guided modeling. The focus is not just training a classifier, but building robust baselines, testing physics-aware structure, and releasing reusable model artifacts.
Highlights
- PyTorch pipelines for gravitational lens classification and finding.
- ROC-AUC based evaluation.
- Physics-guided modeling experiments.
- Public Hugging Face model artifacts.
- Research-style notebooks with reproducible setup notes.
Stack: Python, PyTorch, scientific ML, computer vision, ROC-AUC evaluation, Hugging Face
This workspace explores IBM Research's Mellea generative-programming ideas for structured, testable LLM workflows. I use it to prototype agent loops, MCP demos, evaluator patterns, and more controlled LLM application design.
Focus areas
- Structured LLM calls instead of one-off prompts.
- ReAct-style workflows and tool-use loops.
- MCP-oriented agent demos.
- Evaluation loops for checking LLM behavior.
- Small, inspectable examples that can grow into production workflows.
Stack: Python, Mellea, LLM agents, MCP, evaluators, generative programming
AtomMail is a Hackfest 2nd-prize AI email assistant built around retrieval, OCR, user-history memory, and LLM response generation. It helps users search email context, understand attachments/screenshots, and draft personalized replies.
What it combines
- Retrieval-augmented generation over email context.
- OCR for images and attachments.
- User-history retrieval for personalized responses.
- LLM-based reply drafting.
- Product-oriented workflow for inbox assistance.
Stack: JavaScript, RAG, OCR, LLM APIs, retrieval, email automation
A compact alignment pipeline for preference optimization under limited hardware constraints. The repo explores Direct Preference Optimization with LoRA/QLoRA so alignment experiments can be run without large training infrastructure.
Core ideas
- Preference optimization with DPO.
- Parameter-efficient fine-tuning with LoRA.
- 4-bit quantized loading with QLoRA.
- Modular training and evaluation utilities.
- Hardware-aware experimentation.
Stack: PyTorch, Transformers, PEFT, TRL, bitsandbytes, LoRA, QLoRA, DPO
These repositories are implementation-first study projects. They are intentionally lower-level and help me sharpen fundamentals.
| Repo | What It Covers |
|---|---|
| Mistral Summarization | QLoRA/LoRA fine-tuning for abstractive summarization |
| Language Modeling CS336 | tokenizer, transformer, training-loop, and language-modeling fundamentals |
| DQN Space Invaders | convolutional Deep Q-Network agent for Atari Space Invaders |
| Area | Tools and Concepts |
|---|---|
| ML / DL | PyTorch, TensorFlow, scikit-learn, CNNs, transformers, evaluation metrics |
| LLMs | RAG, LoRA, QLoRA, DPO, summarization, agent workflows, prompt/evaluator loops |
| Backend | FastAPI, Python services, REST APIs, data pipelines |
| Cloud / Data | AWS Bedrock, DynamoDB, S3, boto3, Hugging Face artifacts |
| Frontend | Next.js, React, TypeScript, JavaScript |
| Research Workflow | notebooks, reproducible experiments, metric reporting, artifact release |
- AI/ML internships and research engineering opportunities.
- Applied LLM systems, agentic AI, and ML infrastructure work.
- Collaborations around scientific ML, evaluation, RAG, and model fine-tuning.
- Projects that need both experimental depth and product sense.