AI Engineering Labs

Run the ideas. Inspect the limits.

These are reproducible public artifacts, not hosted live systems. Each card names what can be run locally, what evidence exists, and what remains unproven.

• Market Analyst Agent

  Run locally: A LangGraph research agent with reasoning routes, checkpointed memory, tools, a guardian layer, runtime topology, and harness controls.

  Inspect: How one reference system connects the six parts of the agentic stack.

  Artifact: Source repository · Architecture series

  Limitation: Local reference architecture; no public production reliability benchmark.

• RAG Evals Demo

  Run locally: Evaluation notebooks and modules for retrieval, filters, reranking, faithfulness, LLM judges, and latency on SciFact with embedded Qdrant.

  Bounded result: The repository pins the article’s worked retrieval example at Recall@5 = 0.750, MRR = 0.625, and nDCG@5 = 0.627.

  Artifact: Source repository · Evaluation guide

  Limitation: Worked and SciFact results are corpus-specific; rerun the harness on your own questions and evidence.

• TypeScript Agent Service

  Run locally: A production-shaped TypeScript monorepo with streaming API, strict schemas, tool execution, persistence, worker, MCP, observability, and tests.

  Inspect: The boundary-by-boundary mapping from a Python service stack to TypeScript.

  Artifact: Source repository · Implementation guide

  Limitation: Local service scaffold; deployment and provider calls require your own infrastructure and credentials.

• Search Ranking Stack

  Run locally: BM25, dense retrieval, reciprocal-rank fusion, cross-encoder reranking, and LLM listwise ranking on sampled Amazon ESCI product-search data.

  Inspect: Per-stage relevance, latency, and cost rather than one opaque end score.

  Artifact: Source repository · Staged analysis

  Limitation: Sampled benchmark; the stale LLM row is retained as a warning, not a result.

• Model Compression Demo

  Run locally: Quantization planning, recipes, dry runs, and benchmark configurations for vLLM, LLM Compressor, and TensorRT-LLM workflows.

  Inspect: Which compression path fits a model, target hardware, quality budget, and serving constraint.

  Artifact: Source repository · Quantization guide

  Limitation: The public material supplies plans and recipes; benchmark claims still require the target serving hardware.

Prefer the explanation first? Browse the Work page or read Edge of Context.