AI Lab
Experiments and production patterns across LLM applications, multi-agent workflows, RAG, MCP tooling, semantic search, and vector infrastructure.
A local prompt workbench for shaping solution architecture.
Simulated architecture output
Experiments
Each experiment documents the overview, architecture, implementation snippets, and lessons learned.
A LangGraph supervisor routes inbound support tickets to specialist worker agents, cutting first-response triage time from hours to seconds.
Combined dense vector search and BM25 keyword retrieval inside Postgres, fused with RRF, to fix the exact-match blindness that pure embeddings suffer from.
A golden-dataset eval harness with rubric-based LLM-as-judge scoring, so prompt changes ship on evidence instead of vibes.
A Model Context Protocol server that exposes a CRM's contacts, deals, and activity timeline as typed tools any MCP client can call safely.
An OpenAI Agents SDK agent that handles order lookups, cancellations, and refunds behind layered guardrails — because a refund tool with no ceiling is a career-limiting deploy.
Replaced a brittle Elasticsearch synonym pipeline with Qdrant-backed semantic search over two million SKUs, holding p95 under 80ms with filtered HNSW.
A LangGraph pipeline that drafts, reviews, and routes contract documents — pausing at interrupt points for human sign-off and resuming days later from a Postgres checkpoint.
A TypeScript gateway that normalizes streaming across Anthropic, OpenAI, and Bedrock, with health-scored fallback routing so one provider's bad day isn't ours.
A research spike on A2A-style delegation: opaque agents advertising capability cards, negotiating tasks, and reporting results across trust boundaries.
A reproducible benchmark of three vector stores under realistic enterprise RAG conditions — metadata filters, concurrent writes, and multi-tenancy — not just raw ANN throughput.