Works

NLP / LLM

MCP Walkthrough

An MCP-powered research assistant — from basic tool calling to a multi-server chatbot with resources, prompts, and Claude Desktop integration.

LegoLLM

Ongoing

Complete LLM framework from first principles — tokenization, attention, training, generation, alignment as modular Lego pieces.

LangGraph Research Assistant

Multi-agent research system with parallel interviews, map-reduce synthesis, and human-in-the-loop approval.

Task mAIstro

AI task manager with persistent memory (profile, todos, instructions) deployed via Docker Compose with Postgres and Redis.

RAG from Scratch

Minimal RAG pipeline built from scratch with NumPy — no framework abstractions.

Semantic Search with LangChain

Natural language querying over PDFs using LangChain, ChromaDB, and HuggingFace embeddings.

LangChain RAG Patterns: Agent vs. Chain

Comparing LangChain Agent and Chain RAG patterns with multi-provider support.

LangChain Email Agent

AI email agent with auth-based tool access and human-in-the-loop approval via LangGraph.

                    graph LR
                        A[SmolLM] --> B[LoRA Fine-Tuning]
                        B --> C[AI-as-a-Judge Eval]

SmolLM Fine-tuning

Fine-tuning HuggingFace's SmolLM model on a toy instruction dataset with LoRA and AI-as-a-judge evaluation.

Gemini Long Context

Endangered language translation and codebase security analysis using Gemini 1.5 Pro's long-context window.

LoRA Insights: PEFT Recipes

Reproducing and extending LoRA/QLoRA experiments across Llama 3.2 and Qwen 2.5 on H100 GPU.

PEFT-FLANT5

Parameter Efficient Fine-Tuning of FLAN-T5 model for Dialogue Summarization.

                    graph LR
                        A[FLAN-T5] --> B[PEFT Fine-Tuning]
                        B --> C[PPO Training]
                        C --> D[Detoxified Model]

PEFT-FLANT5-Detoxification

Fine-Tune FLAN-T5 with Reinforcement Learning (PPO) and PEFT to Generate Less-Toxic Summaries

Captcha-OCR

OCR in captchas with Keras

AI Engineering / MLOps

112 tok/s

Phi-3.5 Mini · Q4 · GPU

LLM Inference Benchmarks

Benchmarking llama.cpp, ExLlamaV2, and Ollama for production medical report generation at Lunit.

                    graph TB
                        subgraph Registry ["📦 Model Registry"]
                            M["🤖 Models"]
                            A["📁 Artifacts"]
                            Meta["📊 Metadata"]
                        end
                        style Registry fill:#1a202c,stroke:#9ca3af,color:#d1d5db,stroke-width:2px
                        style M fill:#2d3748,stroke:#9ca3af,color:#e2e8f0
                        style A fill:#2d3748,stroke:#9f7aea,color:#e2e8f0
                        style Meta fill:#2d3748,stroke:#f59e0b,color:#e2e8f0

Lunit Model Registry

A centralized model registry for managing the complete lifecycle of machine learning models.

MLOps Orbit

This project marks the beginning of MLOps journey. Instead of focusing on piece of production-grade machine learning, I focus on building full end-to-end pipeline from data collection to deployment.