Ryan Lazar

Data & AI Engineer

I build data platforms
and the AI agents
that run on them.

From Terraform-provisioned cloud infrastructure to agentic systems that reason over live enterprise data — I work the full stack and ship things you can actually click on. Currently focused on the Oracle data ecosystem.

See the work Resume

Selected work

Flagship — live demo

FleetSentinel

Plant-level telemetry for a hall of liquid-cooled AI compute racks. Streaming anomaly detection (EWMA z-score, CUSUM drift, isolation forest) feeds a deterministic severity engine, a robust trend forecaster predicts when a signal will breach its hard limit — "coolant flow breaches in ~22 min" — and incidents roll up from rack to CDU loop to facility PUE and capacity headroom, inferring shared cooling failures from correlated racks. Claude writes the operator shift briefing; the grading path stays ML-free so every page is explainable.

  • Python
  • scikit-learn
  • Time-series forecasting
  • Claude API
  • FastAPI
  • Podman
Live demo GitHub
FleetSentinel dashboard showing a data-center hall map with rack severity, time-to-breach predictions, and an operator shift briefing

Flagship — live demo

AskMyData

A conversational analytics agent where everything agentic lives inside the database: in-database vector search over supplier contracts, natural-language SQL over live order data, and persistent agent memory in plain SQL tables. Claude reasons in a visible tool-use loop, streamed step-by-step to the browser. Provisioned end-to-end on OCI with Terraform and served over HTTPS.

  • Oracle AI Database 26ai
  • Claude API
  • FastAPI
  • Terraform
  • OCI
Live demo GitHub
AskMyData answering a penalty-claims question: a ranked table of claimable amounts by supplier, with a caveat that most Type A claims are already time-barred under the contract

Autonomous Trader

Autonomous paper-trading system blending momentum signals with congressional-disclosure and news-sentiment following. A deterministic risk scorer auto-executes low-risk trades and routes the rest to a human approval queue — position caps, loss halts, and a kill switch apply regardless. 229 regression tests, each built from a real observed failure.

  • Anthropic API
  • Alpaca
  • Supabase
  • Python
View on GitHub

Investment Monitor

Autonomous two-agent system over the Model Context Protocol: one agent gathers live market data, a deterministic engine grades portfolio risk (beta, Sharpe, concentration), and a second agent writes a validated plain-language daily briefing. Every run audit-logged.

  • MCP
  • Anthropic API
  • Supabase
  • Alpaca
View on GitHub

Investment Tracker

Full-stack portfolio tracker: Next.js frontend with Supabase auth and row-level security, Python microservices for live market data, AI-generated analysis, and professional PDF reporting.

  • Next.js
  • TypeScript
  • FastAPI
  • Supabase
View on GitHub

NYC Taxi Analytics

ETL and analytics over 3.6 million NYC taxi trips: Airflow-orchestrated PySpark jobs feeding PostgreSQL, explored through an interactive dashboard of maps, fares, and trip metrics.

  • Airflow
  • PySpark
  • PostgreSQL
  • Next.js
View on GitHub

About

I'm an engineer focused on the seam between data infrastructure and applied AI. I like building things end to end: provisioning the cloud with code, modeling the data, wiring up the agents, and shipping an interface people can actually use.

Recent work spans plant-level telemetry and failure prediction for AI data-center infrastructure, agentic analytics on Oracle AI Database, multi-agent systems on the Model Context Protocol, and autonomous trading with deterministic risk controls. The common thread: real systems, deployed and demonstrable, not notebooks — and a bias toward keeping the decision path explainable even when the detection path uses ML.