AI Engineering Highlights
- Evals & benchmarks: eval harnesses on golden corpora with a real-user holdout — structural checks gate the run, semantic metrics (P/R/F1, quantity accuracy, cost per run) are tracked as trends, and every production defect becomes an eval case
- Data-driven decisions: picked the import model at 89% vs 76% quantity accuracy for equal cost, rejecting a 9× cheaper model that dropped whole recipes; chose a physics model by error against metered lab values (RMSE 0.117 → 0.081, 404-recipe corpus); made the cheap assistant tier the default after measuring tiers
- AI cost engineering: cost per successful task as the model-selection KPI; micro-USD reserve-and-settle ledger with budget pools; CI spend cut ~70% after attributing 5,343 runs; agent-facing check output ~180K tokens → ~25 lines; MCP tool-result context cut ~75%
- LLM platform, agents & MCP: single LiteLLM egress to Anthropic and OpenAI with server-side routing; bounded agent loops with deterministic tool steps and typed outcomes; public remote MCP server (Go SDK, OAuth 2.1 + PKCE, per-tool scopes, 30 tools, widgets for Claude and ChatGPT)
- AI-native delivery: 2,000+ Claude Code and 190+ Codex sessions orchestrating 15K+ sub-agent runs — parallel worktrees, agent teams, hooks, rules and skills, with review gates on every PR
Experience
- Lead architecture of a multi-cluster Kubernetes platform for 20K+ edge devices (10,000+ RPS, 50+ microservices, 99.9% uptime, SOC2 and GDPR), setting company-wide standards for reliability, IaC and developer experience
- Own DORA and SLO/error-budget KPIs: GitOps with ArgoCD and Policy as Code took deployments from weekly to 50+/day; MTTR cut by 80% through the SRE roadmap and standardized observability
- Showed from git history that ~50 developer changes in 3 months reached production through only 3 manual SRE promotions (16:1); won leadership approval for a multi-tenant Developer Platform where one manifest provisions ~38 resources and new apps go live in about an hour, even for non-developers
- Architected infrastructure for the client's LLM products — on-prem LLM appliance control plane (vLLM, LiteLLM) and MLOps platform (MLflow, agent hosting) on dedicated EKS — across 15 ADRs
- Led coding-agent adoption: production read-only guardrails for agents with a regression suite, agent-readable docs, Claude PR review across 6 repositories, AI-assisted Terraform plan review; mentored 10+ engineers, lately on Claude Code and Codex
- Designed air-gapped and hybrid Kubernetes solutions for regulated industries (government, defense, healthcare), enabling $2M+ in qualified sales pipeline
- Established security scanning (Trivy, Snyk), secrets management (Vault) and FinOps (Kubecost); cut infrastructure spend by 20% ($250K+ annually) with Karpenter, Spot instances and right-sizing
- Built self-service provisioning tooling that reduced ops tickets by 60%
Platform: AWS EKS, Kubernetes, ArgoCD, Terraform, Istio, Prometheus, Grafana, Vault, Go, Python, PostgreSQL, Kafka
- As Solution Architect and Team Lead, led 12 engineers building a real-time media streaming platform for the news industry: 10,000+ RPS with GPU-accelerated live video transcoding, secure private VPC → DMZ content delivery and integrity verification across distributed storage
- Migrated a monolith to containerized microservices, reducing deployment time by 80%
- Implemented CI/CD with Jenkins and GitLab CI and IaC managing 100+ cloud resources
Full-stack development and DevOps across backend, frontend, and infrastructure. Built ETL pipelines processing 1M+ records/day, monitoring systems for 200+ services, and HIPAA-compliant infrastructure.
Notable Projects
Production AI SaaS — confidential (2025-Present)
AI-native B2B SaaS for scientific food formulation behind the highlights above, built end to end with coding agents and live on web and mobile: 4,100+ commits, 3,100+ merged PRs, 10K+ Go tests in the first ten months.
AI-assisted Terraform Plan Review (2026)
Opt-in LLM review of Terraform plans in a Go Terragrunt CLI for the SRE team — masked, allowlisted export to a local Claude Code or Codex CLI; advisory only.
Agent Trace Cost Estimator (2026)
Deterministic parsing plus cheap-model labeling of agent-trace digests into a token and cost estimation catalog (963 sessions, ~57B tokens).
Disaster Recovery, Migration & Load Testing (2021-Present)
DR for 5 critical data stores with RPO <30min and RTO <2h; migration of 20+ customer clusters to modern Kubernetes with zero data loss; load testing and chaos framework validating 10,000+ RPS before production.