Streaming civic-data pipeline: Chicago 311 → Kafka → Claude urgency classification (Langfuse-traced) → DuckDB or Snowflake via idempotent MERGE → dbt SLA marts. Runs with no accounts: make demo-local.
Demonstrates: Streaming LLM classification into a warehouse with idempotent MERGEs, a DLQ, and dbt tests that fail on drift.
See what a company newly disclosed, quietly dropped, or materially reworded between filings — with every claim traceable to the exact span it came from.
Demonstrates: Separating deterministic diffing from model judgment, then scoring the model half against labels.
A cleaned, versioned, queryable view of public hospital price and provider data — with the reconciliation failures published rather than hidden. CMS price transparency + NPPES + Synthea FHIR.
Demonstrates: Publishing what the sources got wrong, and gating the pipeline so a run that fails cannot ship anything.
Read-only MCP server for hospital price questions: six typed tools over clear-pricer's pinned, hash-verified release of three Chicago hospitals' price files and the NPPES registry, every row cited to the hospital's own file.
Demonstrates: Exposing a versioned data release as typed, grain-safe MCP tools, verified by planted-bug mutations.
Six Model Context Protocol servers over curated biomedical data — ClinicalTrials.gov, openFDA (labels/events/recalls/approvals) and PubMed — pairing a pgvector hybrid-search document store with a gated relational staging layer, cited LLM tools, cross-source SQL joins, incremental refresh and usage metering.
Demonstrates: Retrieval over a document store with versioned relational staging under it, each gated and published.
A reviewer queue for payment-integrity claims on synthetic data. Three deterministic rules flag duplicate billing, unbundling and out-of-network adjudication; every flag shows the fields that fired it and a grounded explanation; a human approves, dismisses or escalates. Live demo runs cached, no API key.
Demonstrates: Rules detect, an LLM explains, a human decides — the rules also in SQL, proven equal, then two recall gaps closed.
A governed AI tool hub that removes the prompt — 16 cited-or-abstain tools across 4 platforms on one no-prompt shell, safe on sensitive data (PII/PHI sanitized before egress), with a measured 20/20 trust eval (0 hallucinations). Live demo included.
Demonstrates: Grounded generation with cite-or-abstain, measured on a labelled set and on a harder one whose misses are published.
Score a grounded-answer (RAG) system on citation accuracy, abstention, hallucination, and cost and latency — offline, with an audit log a compliance reviewer can read, and a CI gate that fails a PR on a regression.
Demonstrates: Extracting a proven method into an installable tool, and using it to find a real gap in a shipping system.
Test LLM systems like software — grade against golden + adversarial sets (LLM-as-judge + deterministic rule checks), catch regressions in CI, and filter PII/hallucinations. Validates itself: the judge is calibrated against a human gold set.
Demonstrates: Validating the evaluator itself — a judge calibrated against human labels, and an injected break proven caught.
A daily Airflow and PySpark pipeline that pulls NBA box scores into partitioned Parquet and a model-ready feature table. Ran unattended through the 2025-26 playoffs and Finals against a source that changed underneath it. 5-task DAG, 139 tests, live Streamlit dashboard.
Demonstrates: Leak-free walk-forward evaluation, reported against three named baselines even when the model loses.
A delivery artifact for AI opportunity management: structured intake and explainable five-dimension scoring, and — the part that matters — it flags where AI is a poor fit and ranks those cases separately, so a bad idea cannot outrank a good one.
Demonstrates: Structured AI-opportunity scoring a human can override, with a first-class judgment about where AI does not belong.
The planner and step-cost layer: it turns a goal into an inspectable plan and runs it under guardrails — a step budget, retry-then-halt, a human checkpoint, schema-validated output, an allow-listed registry. Project Suver is the governed tool surface those steps would call.
Demonstrates: Running an agent under enforced guardrails, with every guardrail exercised and its outcome published.