Information Retrieval
Black-box retrieval evaluation harness
Built eval_retrieval.py to empirically compare three retrieval strategies — semantic-only search, hybrid BM25 + dense retrieval with Reciprocal Rank Fusion, and hybrid retrieval with cross-encoder reranking — against a shared query set. The evaluation confirmed the reranker's accuracy gains came at a memory cost that exceeded a 512MB free-tier hosting budget, turning a modeling question into a documented production tradeoff.
Agentic Systems
Multi-tool research agent (LangGraph + MCP)
Developing a research agent using LangGraph's create_react_agent, tool access via the Model Context Protocol (through langchain-mcp-adapters), Groq for inference, Tavily for search, Qdrant for retrieval, and LangSmith for run-level observability and debugging.
Model Adaptation
LoRA fine-tuning experiments
Part of a structured 3-week applied AI engineering plan: parameter-efficient fine-tuning with Hugging Face PEFT, paired with Redis caching, Docker containerization, GitHub Actions CI/CD, and quantization work on the EchoSign sign-language model.
In progress