Proven impact
Quantified results from production cloud-native platform engineering.
What my tooling actually outputs
Anyone can list technologies. This is a real run of Forward-Deployed Preflight against a customer environment — every blocker gathered in one pass, each with the specific remedy and the team that owns it.
$ python main.py --profile examples/acme-staging.yaml --skip-tags k8s,llm Preflight — acme-staging ======================== [FAIL] dns 2 of 3 required hostname(s) do not resolve -> Add these names to the resolver used by the deployment target (not just the engineer's laptop — split-horizon DNS resolves differently per network segment). If they are public names, check whether egress DNS is restricted to an internal forwarder: registry.acme-internal.net, postgres.staging.acme-internal.net [FAIL] tcp 2 of 3 endpoint(s) unreachable -> Request an egress allow-rule for these host:port pairs from the network team, from the subnet the workload will actually run in. A 'connection refused' means something answered and declined (likely the right path, wrong port); a timeout means the packets were dropped (firewall). [ok ] tls_interception TLS to api.anthropic.com presents a recognised public issuer [ok ] http_egress egress works (https://api.anthropic.com/v1/models returned HTTP 401) 2 passed · 0 warning(s) · 2 failure(s) · 0 could not run NO-GO — resolve the above first
Two details matter more than the checks themselves. Every finding carries its own fix — a report that just says FAIL: unreachable has pushed the diagnosis back onto the customer, so the code raises if an actionable finding has no remedy. And "could not run" is counted separately from "passed": a check that couldn't execute reports BLOCKED, which outranks FAIL, because a green report that means "we didn't look" gets discovered in the customer's deployment window. The same rule runs through every project below: a gate that cannot fail is not a gate.
How I can help
Focused engagements with clear outcomes and a defined finish line. Scope and day rate — let's discuss on a call.
Kubernetes & Cloud Platform Build
Production-grade clusters on AKS / GKE / EKS, provisioned as code.
- Terraform + Helm, multi-environment
- Ingress, autoscaling (HPA), observability
- RBAC, network policies, secrets management
DevSecOps Pipeline Hardening
Shift-left security baked into your delivery pipeline.
- SAST / SCA / secret & image scanning
- GitOps with ArgoCD, safe rollouts & rollback
- Policy-as-code & supply-chain security
AI Infrastructure & LLMOps
Deploy and secure LLM / agentic-AI workloads at scale.
- LLM inference serving & RAG on Kubernetes
- MCP servers & multi-agent automation
- AI-security guardrails (PII, injection, cost)
Forward-Deployed / Solutions Engineering
Getting your product live inside a customer environment you don't control.
- Pre-deployment environment verification
- Egress, TLS interception & least-privilege RBAC
- Pilot readiness: model access, rate limits, data residency
Agent Engineering
Multi-agent systems that do real work and can be trusted with it.
- LangGraph pipelines with tool use & state
- Hard cost caps, dry-run modes, audit trails
- Evaluation harnesses — not vibes
Cloud Cost & FinOps
Find the spend, cut it, then keep it from creeping back.
- Multi-cloud cost attribution & anomaly alerts
- Rightsizing, autoscaling & spot strategy
- 40% reduction delivered in production
Selected work
Open-source, production-style platforms — each with architecture docs, CI, offline tests, and one-command setup. Filter by what you need built.
How I work
Four rules I apply to every engagement. They come from production failures, not from a blog post.
A gate that cannot fail is not a gate
Most "monitored" systems fail silently because the check was wired so it could only pass. I verify every guard by planting a violation and watching it fire — before I trust it in production.
Uptime is not a result
A pipeline that exits green on every run while producing nothing is the most expensive kind of broken. I report outcomes, name the bottleneck stage, and make activity metrics secondary.
A wrong report is worse than a missing one
Missing data makes you look. Wrong data makes you act — and debug five healthy components. Diagnostics get the same scrutiny as the code they describe.
Ship in reviewable increments
Small PRs with green CI, tests that state the defect they prevent, and commit messages that explain why. You stay in control of what lands and can hand it to your own team afterwards.
Tech stack
Cloud-native, security, and AI-infrastructure tooling I work with daily.
Cloud & Kubernetes
CI/CD & Security
AI & Agents
Observability & Data
Let's build something secure
Tell me what's broken or what you're trying to ship. If I'm not the right person I'll say so on the call rather than after the invoice.
Currently taking on new contract work — London, UK or fully remote