Projects

Multi-Agent Orchestration Framework

A framework for coordinating specialized LLM agents (planner, executor, critic) with shared memory and tool access, for tasks too complex for a single prompt-response loop.

Long-horizon tasks break down when a single agent tries to plan, act, and self-correct in one pass. This project explores an orchestrator pattern that splits those responsibilities across specialized agents with a shared state store, so each step’s context stays focused and failures are recoverable without restarting the whole run.

Source

Retrieval-Augmented Generation Pipeline

A RAG pipeline combining hybrid search and reranking to ground generative model output in a client's proprietary knowledge base, with citations back to source documents.

Generic vector search alone misses exact-match cases like product codes or error messages. This pipeline pairs dense embedding search with keyword (BM25) retrieval, reranks the merged candidates, and forces the generation step to cite the passages it used — cutting down on unsupported answers.

LLM Agent Evaluation & Observability Dashboard

An analytics dashboard tracking agent trace quality, latency, cost, and hallucination rate across runs, used to catch regressions before they reach production.

Agentic systems fail in ways unit tests don’t catch — a tool call that technically succeeds but returns the wrong thing, or a plan that loops. This dashboard turns agent traces into metrics (task success rate, steps per run, cost per completion, LLM-as-judge scores) so regressions show up as a chart moving, not a support ticket.

Autonomous Coding Agent Prototype

An agent that plans, edits, and tests code changes in a sandboxed repository loop — reading its own test/build output to decide the next action instead of following a fixed script.

Built to explore how far a plan → act → observe loop can get on real software tasks: given an issue description, the agent reads the relevant files, makes an edit, runs the test suite in a sandbox, and reacts to the result — self-correcting instead of stopping at the first failure.