Every metric in my AI systems has a receipt — a committed JSON in the repo you can clone and re-run to get the same number.

This year I shipped 6 open-source AI systems — each with tests and reproducible results, not demos and vibes. The flagship is Atlas AI Platform, a hexagonal, 12-factor AI-agent platform: policy-based model router, hybrid retrieval, layered guardrails, a grounding critic, and trace + cost observability.

What I do

• Multi-agent orchestration — LangGraph: routing, tool use, grounding, retries
• Hybrid RAG — dense + BM25 + SQL, RRF fusion, reranking, citation tracking, evaluation
• AI guardrails & security — prompt-injection detection, PII redaction, safe tool execution
• LLM evaluation & observability — LLM-as-judge, groundedness, latency & cost receipts
• ML platform foundation — Databricks, MLflow, Delta Lake, Azure

The thing most engineers skip: eval rigor and honesty. Most public AI demos hide their methodology — mine document the labeling protocol, disclose the caveats, and ship the raw reports, including the times a simpler baseline won.

12 years inside freight/logistics and financial data systems — finding the kind of data inconsistencies AI is now being asked to catch. I've seen the messy real-world data most AI demos avoid; my systems handle it.

◆ Open To — currently in the Detroit metro area; home base Greater Toronto Area. TN visa eligible (US employers sponsor via a straightforward process, not the H-1B lottery). Open to AI/ML Engineer, Applied AI, and Forward-Deployed Engineer roles — Toronto, remote-US, or US relocation.

Languages

Interfaces & Services

Backend & Databases

Cloud, DevOps & Tooling

AI / ML / Data

October 2021 – Present · 3PL / Freight & Logistics

Building production AI systems and MLOps infrastructure on top of multi-year data engineering work. Recent focus: multi-agent orchestration, hybrid RAG, LLM evaluation, and AI guardrails.

Independent open-source AI engineering (personal projects, built outside work — every number has a committed receipt):

• Atlas AI Platform — hexagonal 12-factor AI-agent platform: FastAPI gateway (auth, rate limiting, role-scoped endpoints), policy-based model router with fallback, hybrid retrieval (dense + BM25 RRF + guarded SQL), layered guardrails, grounding-critic orchestration, and OTel-style trace + cost observability. 57 tests, 8 ADRs, Docker + CI
• freight-rag-agents — multi-agent hybrid RAG on LangGraph with a grounding critic (routing 95% / retrieval 100% / all 3 groundedness traps caught, 20-query labeled eval)
• databricks-mcp — MCP server giving agents safe read-only SQL, with an sqlglot AST guard blocking DDL/DML and filesystem exfiltration
• production-rag-eval — reproducible RAG eval (P@5 0.68 / MRR 0.92) + a multi-method benchmark; reported honestly that dense retrieval beat hybrid/GraphRAG/reranking
• mlx-lora-finetune — LoRA fine-tuning of Qwen2.5 (4-bit) via Apple MLX: 71% → 100% field accuracy on structured freight extraction, ~6 min on an M4
• voice-agent — voice pipeline with local STT (faster-whisper → tool-calling GPT-4o-mini → TTS): 3.18s p50 end-to-end, per-stage receipts

Data & ML platform engineering:

• Built Databricks + PySpark pipelines over multi-million-record freight datasets, improving batch performance 30%+
• Built data ingestion with Azure Data Factory and set up MLflow tracking + registry; developed internal APIs integrating analytics with operations and finance systems

Python LangGraph FastAPI MCP sqlglot pgvector DuckDB Databricks MLflow Azure Docker

January 2017 – October 2021 · Brampton, ON

Built the data engineering foundation that powers current AI/ML work — owned data infrastructure for a multi-million-dollar freight logistics business: ingestion, transformation, reconciliation, reporting.

• Built billing-reconciliation systems that surfaced data inconsistencies across multiple carrier sources
• Automated data validation, reconciliation, and reporting workflows, reducing manual effort and processing delays
• Self-taught Python, SQL, and cloud platforms while delivering production value — set the trajectory for current AI engineering work

SQL Python ETL Data Reconciliation Financial Data

October 2013 – December 2016 · Brampton, ON

Freight movement, billing, and reconciliation analytics for a 3PL business. Drove the case for moving from spreadsheet workflows to proper data infrastructure.

• Managed and analyzed shipment, billing, and logistics data across multiple carriers
• Built deep operational knowledge of end-to-end logistics, freight reconciliation, and 3PL supply chain — domain expertise that now informs my freight-focused AI systems

Microsoft Azure

Databricks

Google

PMI

building:
  - Atlas AI Platform            # hexagonal AI-agent platform -> Phase 2 (workflow engine, more workloads)
  - Multi-agent orchestration    # LangGraph routing, grounding critics, retries
  - AI guardrails & security     # prompt-injection, PII, safe tool/SQL execution
exploring:
  - Cross-encoder reranking and where hybrid retrieval actually pays off
  - LLM-as-judge reliability & graded relevance
  - Distributed / async agent execution (Kafka, Ray) as the platform scales
open_to:
  - AI Engineer · Applied AI · Forward-Deployed Engineer · ML Platform
  - Toronto · Remote-US · US relocation — TN visa eligible

"Build for production: reproducible, measured, and honest about the trade-offs."