Leadership
A career spent leading engineering and AI across very different organizational contexts: from federal policy to Big Tech, from a PhD lab to running AI & ML at a venture-stage startup and founding an AI-native product on the side. Below is the arc with scope, ownership, and outcomes for each role.
Running the applied-AI function at a venture-stage AI company.
Leading a team of AI/ML scientists delivering an end-to-end agentic AI platform (search, conversation, and tool orchestration), scaling the function alongside company growth. Full-stack AI leadership: roadmap, team, architecture, eval, and the external technical narrative.
Scope
Applied-AI function end-to-end: AI strategy, OKRs, hiring and performance, model and infra decisions, responsible-AI posture, and the technical story we take to customers, partners, and investors.
What I own
- Architected a multi-turn Conversational AI engine with agent memory, planning, and tool-use: the primitives behind the platform's agentic behaviour.
- Shipped a hybrid embedding + LLM taxonomy system hitting 99% accuracy at <600 ms latency across millions of SKUs.
- Built a self-evolving AI Taxonomy Generator and a multi-modal enrichment pipeline that cut manual maintenance by >90%.
- Designed a closed-loop ingestion system connecting live user traffic to auto-evaluation, retraining, and fine-tuning.
- Partnered with product, engineering, and GTM leadership to define the AI roadmap and OKRs across stakeholder tiers.
- Set the hiring bar and interview loop for applied-science hires.
Outcome
End-to-end ownership from model design to production. Building the AI platform from the inside out and raising the engineering bar on what "shipping agentic products" means.
Founder of a vertical AI product for endurance athletes.
Designed and built Athlete Space end-to-end: a typed multi-agent orchestration layer on top of a modern LLM stack, grounded in CTL/ATL/TSB training-load science. I own the technical architecture, the product thesis, and the investor narrative.
Scope
Full stack: domain model, planner pipeline, coach orchestrator, evaluation loop, product, and fundraising.
What I own
- V5 specialist-agent architecture (planner, advisor, evaluator, verification, reply synthesiser, more).
- Deterministic training-load pipeline (macro plan → philosophy → week structure → volume → templates → session text → calendar).
- Python 3.12 / FastAPI / PostgreSQL backend; MCP servers for DB and filesystem.
- Go-to-market thesis and investor conversations.
Outcome
V5 live, regression suite green, rollback plan tested. Real users. See the Athlete Space page.
Applied ML at Amazon scale.
Four years inside Amazon's applied-science org. Led development and deployment of production ML systems for real-time, enterprise-scale decision-making, operationalizing models from prototype to live with reliability and monitoring SLAs.
Scope
End-to-end ML workflows: problem framing, modeling, deployment, and live measurement against a business metric.
What I owned
- ML workflows emphasizing behavioural signal modeling, interpretability, and live monitoring.
- Building systems that learn from real user behaviour rather than offline proxies.
- Cross-functional delivery with product, engineering, and compliance, shipping auditable, responsible ML.
- Reducing infrastructure cost while holding the quality bar.
What I took away
How applied ML actually converts into business value at scale: the engineering hygiene, the review culture, the disciplined separation between a model that's "good" and a model that ships.
Research residency at Google.
Designed production ML pipelines with an emphasis on reliability, low latency, privacy, and Responsible AI in large-scale deployed systems.
What I took away
The engineering discipline behind deploying ML at Google scale: what it takes to hold latency, privacy, and auditability as first-class constraints, not afterthoughts.
Ph.D. in computer vision under Prof. Andreas Savakis.
Four years of full-time research on human pose estimation in RIT's Computer Vision Laboratory. Published 20+ peer-reviewed papers with 500+ citations and a patent in applied AI. Graduated with a 3.96 GPA.
Focus
Scalable architectures bridging theory and deployable systems: multi-scale context aggregation (WASP) applied to single-stage pose estimation.
What I owned
- Full research pipeline: problem statement, architecture design, experiments, writing.
- Lead author on UniPose (CVPR 2020), UniPose+ (IEEE TPAMI 2021), and subsequent extensions.
- Led and mentored graduate research teams; managed sponsored projects end-to-end.
Why it still matters
The architectural taste I developed in the PhD (how to reason about multi-scale context, how to design single-stage systems that generalize) continues to shape how I build production AI systems today.
Applied ML inside the Canadian federal government.
Built predictive risk models on large aviation-safety datasets for UAS collision-scenario simulation, an early foundation in simulation-based decision-support systems, and a first taste of what it takes to ship ML inside a regulated organization.
Scope
Applied ML serving regulatory decisions; stakeholders were policy and operational staff, not engineers.
What I took away
How to translate between the model and the decision. It's where I learned that models only matter when they clear a real decision.