I build AI systems that survive production. Agentic platforms, retrieval, and the security and reliability engineering that keeps them alive when the environment changes.
Most people shipping LLM apps stop at "it works." I'm interested in everything that happens after.
Production AI that takes real actions is an attack surface, a reliability problem, and a moving target — all at once. That's the part I build for. I took an agentic platform from first commit to org-wide deployment solo, then kept it alive: red-teamed it, watched its environment shift under it, and engineered it to heal itself.
Underneath the AI sits eleven years of infrastructure at scale — multi-cloud estates, network architecture across 150+ sites, HIPAA-regulated environments, IT operations for organizations with thousands of employees. I know what survives contact with production and what only works in the demo.
Software that reasons and acts is an attack surface, not a demo.
A deployment is done when it survives its environment changing.
Every failure becomes a regression test — or it happens again.
An agentic AI support platform for a 65-location, 2,000-employee restaurant group. It doesn't just answer — it creates tickets, provisions access, and runs onboarding, in English and Spanish. Designed, built, secured, and operated solo: concept to org-wide rollout in about four months.
A Gemini-powered agent orchestrating 17 least-privilege tools, each an isolated sub-workflow with its own scoped credential. The retrieval layer is built from primitives, not a framework: PostgreSQL + pgvector, task-typed 768-dimension embeddings, multi-query rewriting with rerank and dedup, confidence thresholds calibrated from sampled score data.
Because Al3x takes real actions, I mapped it to the OWASP Top 10 for LLM Applications — 10/10 addressed — then ran an internal red-team against my own system and remediated 16 findings: prompt injection, impersonation, cross-user data access, capability disclosure, mass-exfiltration attempts.
Identity doesn't live in the prompt. It lives on the server, where the model can't touch it.
That one primitive — authenticated identity supplied as fixed context the LLM cannot populate or override — closed the entire IDOR and spoofing class. On top of it: per-object ownership checks, runtime injection and flood detection with escalate-to-block, and Google OIDC verification on every inbound request, rolled out audit-mode-first.
Four days after a zero-downtime migration to Cloud SQL, every database call from production started timing out. The agent kept talking — it just couldn't remember anything.
The platform's cloud runtime publishes no static egress IP — it rotates without notice. The only address allowlisted on the database was my laptop's. Everything worked in testing because I was the one testing.
The obvious fix — allowlist the new IP — just schedules the next outage. And the second failure surfaced immediately: a per-instance TLS certificate the client stack structurally couldn't verify.
A watchdog that checks its own egress IP every five minutes and reconciles the database allowlist through the cloud admin API. It keeps state outside the database — the healer can't depend on the thing it heals — remediates through a path that isn't IP-restricted, and emails me on every change. Self-healing, never silent.
IP rotations now self-heal in under five minutes with zero manual steps, at zero recurring cost — and the allowlist stayed tight instead of being opened up. The residual TLS risk is documented honestly in a runbook, with the full fix evaluated and consciously deferred.
A deployment isn't done when it works. It's done when it survives its environment changing.
Al3x improves itself on a closed loop: failed conversations are captured, replayed in a sandbox against candidate fixes, verified by a deterministic harness independent of the model, and shipped back through a human gate.
Rollout ran under real governance — an ITIL change record with risk matrix, validation evidence, and a backout plan, taken through CAB approval. Enterprise discipline, one-person team.
Systems I've designed, built, and shipped end to end — multi-tenant SaaS, bilingual platforms, self-running automation. Different problems, same bar: production-grade, secured, documented.
A digital business-card platform built as real multi-tenant SaaS: NFC tap and QR scan tracking with per-card analytics, member accounts with strict cross-tenant isolation enforced by Postgres row-level security, rate-limited lead capture, CSP and security headers, and AI-assisted brand theming behind an SSRF-guarded extractor.
A fully bilingual production site where the Spanish is written natively — and a robot enforces it: an automated i18n parity gate fails the build on any English leakage into Spanish routes. Instant-quote estimator, per-locale SEO with hreflang and a bilingual sitemap, lead pipeline wired to n8n + Resend.
A 53-page static site built to prove craft: a three.js contour-field hero, GSAP scroll scenes, self-hosted subset fonts at 113KB total, and a hardened contact pipeline with honeypot and rate limiting. The site is the portfolio.
An intake funnel that runs itself: website webhook → CRM persistence → auto-drafted follow-ups. Designed persist-first — state is durable before any notification fires, so a dead downstream integration can never lose a lead. It already survived one: credentials died silently, the pipeline kept every record.
// all four designed, built, and operated end to end by one engineer
From IT operations to production AI. The common thread: take messy, reactive systems and make them documented, proactive, and — lately — intelligent.
The AI and security stack I build with today, the product engineering around it, and the infrastructure foundation underneath.
Agentic systems · Gemini · RAG from primitives (pgvector, embeddings, rerank) · evals & regression harnesses · prompt engineering · n8n orchestration
OWASP LLM Top 10 · red-teaming · prompt-injection defense · server-side identity binding · least-privilege tool design · OIDC verification
Next.js 15/16 · React 19 · TypeScript · Tailwind · next-intl i18n · Supabase with RLS · multi-tenant SaaS · SEO
GCP (Cloud SQL, Cloud Run, IAM, Secret Manager) · AWS · Azure · Docker · self-healing infrastructure · Windows Server / AD · VMware
PostgreSQL · SQL · Python · REST APIs · zero-downtime migration · persist-first pipelines · Freshservice / ServiceNow / Toast / Resend integrations
Fortinet · Ubiquiti · Cradlepoint 5G · SD-WAN · ITIL / CAB · incident response · KPI/SLA · HIPAA
If you're hiring an AI engineer who ships production LLM systems, secures them, and keeps them alive after launch — let's talk.
daniel_lewis11@outlook.com