AI Agent Developer · Geotab · Toronto
I build AI agents that run in production.
Sole engineer behind Geotab’s multi‑agent platform: a root agent and specialised sub‑agents on Google ADK and Vertex AI, retrieval that cites its sources, deterministic validators around every model call, and tracing on everything. Used across the company by 3,400 people.
Production systems first, then the projects that taught me the most.
Geotab · production · 2026
A root agent and sub‑agents on Google ADK, deployed on Vertex AI Agent Engine. Retrieval with citations over a governed corpus of ~300 documents, typed tools into BigQuery, Google Maps and the HR system, and self‑hosted Langfuse tracing on every call. I own the architecture, the code and the pager.
Geotab · production · 2026
A sub‑agent drafts travel letters, then nine deterministic checks run in parallel with hard‑fail / review tiering. The model only does the semantic work. Passed a formal threat‑model review.
York University · robotics · 2025
YOLOv8‑Pose keypoints → state machine → Cartesian velocity over ROS2. Wave, and a 7‑DOF arm follows.
Geotab · 2025
New developers kept asking “who can access this dataset?” A Google ADK agent answers first.
York University · embedded · 2025
Raspberry Pi + DHT11, validated CSV logging, PySpark Structured Streaming emailing alerts.
Geotab · data enablement · 2024–25
Metadata ingestion with the Python SDK and REST API, giving end‑to‑end lineage from PostgreSQL through BigQuery to Superset. Pipelines across production and UAT.
SPIE proceedings · 2023
Research with Nick Feamster on a 5,172‑message dataset; the improved SVM reached 98.17% accuracy.
What the platform actually runs on.
Write‑ups on agents, validation, and the odd robot.
A webcam, YOLOv8-Pose, a small state machine, and Cartesian velocity commands over ROS2. Wave, and a 7-DOF arm follows.
New developers kept interrupting senior ones with "who can access this dataset?". A Google ADK agent now answers first.
A DHT11, a Pi, validated CSV logging, and PySpark Structured Streaming watching the folder to send alerts. Embedded fundamentals with a data pipeline bolted on.
Toronto, mostly. Friends, teams, stages.
Hiring conversations, agent architecture, or comparing notes on evals. Thirty minutes is easy to find.