Principal Software Engineer · Applied AI & Cloud Platforms
Multi-tenant AI SaaS, inference economics and cloud infrastructure — production systems, not demos. 15+ years shipping end to end.
15+
Years end to end
100k+
Users served
99.5%
Platform uptime
80+
App vendors on my pipelines
Provider-abstracted inference, a gateway your teams call instead of ten SDKs, cost governance with a real kill-switch, and observability that catches silent retrieval failures before your users do.
Tenant isolation you can prove, not hope for — and the billing spine that turns AI usage into revenue: metering, overage, trial-to-paid lifecycle, webhook reconciliation.
Infrastructure as code on AWS, keyless CI/CD, staging as a required gate, migrations that cannot race. Deployment stops being the risky part of your week.
Principal engineer for a digital-takeoff SaaS on AWS Canada Central. I own the backend and the cloud beneath it.
Python / FastAPI on Lambda, multi-stack AWS CDK
Multi-tenant Cognito security — JWT audience isolation so cross-tenant escalation is impossible by construction
Stripe free / trial / paid lifecycle with AI usage metering, overage billing and reconciliation
Document AI (batch OCR / vision, LiteLLM) with a fail-closed cost kill-switch for the worker fleet
Keyless GitHub OIDC deploys with staging as a required gate
GenAI infrastructure: provider-abstracted inference, RAG and MCP implementations, and a unified model-and-tools gateway.
Multi-GPU inference serving on vLLM / Kubernetes — ~25% cost and ~40% latency reduction on that platform
One provider switch across local Ollama and a LiteLLM proxy — identical generation code runs free-local or managed-cloud
Optimized image-manipulation pipelines serving 80+ app vendors
Snowflake AI agents with granular access control and data-residency compliance
Multi-agent orchestration and human-in-the-loop review UIs, productionized with IaC and CI/CD
Canada's leading digital mental-health service, delivering EFAP and CBT to 100k+ employees. I inherited it abandoned and undocumented.
Uptime after the rebuild99.5%
Users on OAuth 2.0 PKCE SSO with Microsoft Identity25k+
Partner organizations federated over SAML 2.05+
Deploy time, down from four hours20 min
DigitalOcean → AWS migration on TerraformZero downtime
Idempotency is a reflex, not a retro item — a platform-wide contract across the whole Lambda fleet, advisory-locked migrations, poison-message handling. Every postmortem ends as a CI-enforced guard, so the same failure cannot recur quietly.
Least-privilege IAM, rotation-aware caching, fail-closed gates, no stored keys in CI. Tenant boundaries are chosen so escalation is structurally impossible rather than merely unlikely — and the same discipline shows up in unpaid side projects.
I build to learn a domain, then delete what an off-the-shelf tool already does — one bespoke 40-file engine came out for a maintained dependency. Less surface area for your team to carry after I hand it over.
Open to principal & staff roles · and consulting
Correctness, cost or observability — that's the conversation worth having.