Projects

RAG Pipeline Architecture

Built a distributed, asynchronous RAG pipeline using Celery, Redis, and Unstructured for document ingestion, with independently retryable stages and exponential backoff for fault isolation. Used Amazon S3 for artifact storage and Qdrant for vector retrieval, enabling independent scaling across ingestion, storage, and search. Implemented a hybrid retrieval service combining vector search, keyword search, metadata filtering, and cross-encoder reranking to improve context quality before LLM generation.

Celery Redis Qdrant Amazon S3 Unstructured

GitHub

Autonomous Bug-Fix Agent

Designed an autonomous LLM agent using LangGraph that diagnoses production crash logs, patches source code, and iteratively re-runs a test suite until verification passes. Built a sandboxed tool-execution layer enforcing path-traversal, symlink-escape, and file-extension validation on every LLM-issued file operation, with atomic writes for safe autonomous modification. Validated the sandbox with a 31-case pytest suite covering 6 attack vectors, catching a silent PATH-related failure in the test runner.

LangGraph Python Pytest

GitHub

Legal Clause Explainer — QLoRA Fine-Tuning

Fine-tuned Llama-3B with QLoRA to generate plain-English explanations of legal contract clauses, converting 80k raw clauses into structured instruction-response pairs through a custom preprocessing pipeline. Built an evaluation pipeline comparing base vs. fine-tuned outputs using BERTScore and custom format-consistency scoring on a held-out test set. Tracked training experiments across multiple LoRA configurations with CometML and deployed the fine-tuned adapter behind a FastAPI endpoint with a Streamlit interface.

Llama-3B QLoRA CometML FastAPI Streamlit

GitHub