AI infrastructure engineer.
Ex-Google SRE.
I’m Benjamin Vigil. I build AI systems and the secure, observable infrastructure underneath them, from Google-scale reliability to cleared DevSecOps work.
Search the record
A small shell over my work. Type a command, or pick one. Every line comes from real work.
Also works: grep vggt, grep stripe, grep rhel, clear.
What I do
I spent three years at Google keeping systems up: the distributed build service behind Google’s source code, and Firebase Cloud Messaging, which delivers push to billions of devices.
Since then I have hardened Kubernetes for Department of Defense IL2/IL4 tenants, run GPU clusters for a generative image service, and now hold a Top Secret DevSecOps role I can’t describe here beyond its stack: RHEL, Red Hat Satellite 6 and Kubernetes.
On my own time I design, build and operate AI products end to end. One is a spatial engine that turns real-estate listing photos into walkable 3D homes in the browser. The other is a platform with paying customers: a network gateway, web dashboards, native iPhone and Mac apps, and real-time monitors.
I work from Colorado Springs, mostly in a terminal, usually beside a fleet of AI coding agents that I direct and review.
Selected systems
About 1,270 commits across ten repositories since June 2026, all mine. Open a row for the stack and the proof.
Spatial AI
-
Turns listing photos, 8192×4096 panoramas and floor plans into one registered, walkable house that runs in a normal browser tab.
Stack Python, PyTorch, COLMAP, ALIKED + LightGlue, VGGT multi-view geometry, Sim(3) pose-graph refinement, Three.js
Proof 424 of 424 route stops, 31 of 31 safe spawns, zero teleports, about 120 FPS in a 12.8 MB runtime
-
Pinned, hash-verified models run on the machine: depth, surface semantics and CLIP scene labels. No source image ever leaves for an API.
Stack Depth Anything, UperNet, CLIP via quantized ONNX, Apple Metal (MPS), strict checkpoint loading
Proof zero API calls per run; every model hash is part of the cache key
-
Trains a 3D Gaussian splat for each scene from registered cameras, as a photoreal layer beside the editable geometry.
Stack 3DGS on Apple Metal, registered camera export, Unreal and Blender pipelines for lighting and materials
Proof kept separate from canonical geometry, so it can never move a registered camera
-
A persistent worker that reasons room by room, builds and reviews, with frozen evidence, retries, hard spending caps and explicit unresolved states.
Stack Frontier LLMs, bounded reason, build and review jobs, World Package v0.1 (geometry, materials, collision, navigation)
Proof a run stopped itself at review at $4.77 of a $5.00 cap instead of shipping a guess
-
A SwiftUI app that guides an ordinary iPhone through a structured property capture for the engine.
Stack SwiftUI, native pose capture, TestFlight
Proof version 0.2 submitted to TestFlight
Production platform
-
A credential gateway that sits in front of mobile and ISP proxy capacity: per-customer auth, per-IP limits, canary rollouts and workers that roll without dropping live tunnels.
Stack Node, HAProxy, systemd worker templates, OVH bare metal
Proof private ranges blocked at the socket (SSRF), hashed auth buckets, admission control under load
-
Checkout, dashboards, native API, transactional email, Discord integration and SEO, all serving real paying customers.
Stack Next.js 15, React 19, TypeScript, Prisma, Postgres (Neon), NextAuth, Stripe with automatic tax, Resend, Vercel
Proof 576 commits; customer-safe DTOs so supplier data never reaches a browser
-
Customer and staff apps over versioned mobile APIs, with Stripe-hosted checkout, push alerts and blocking legal acceptance.
Stack Expo, React Native, TanStack Query, Zod, APNs
Proof memory-only access tokens, rotating Keychain refresh tokens, HTTP 428 legal gate, 57-person TestFlight beta
-
One identity across isolated provider workspaces, with billing, payouts, reviews, a live activity feed and an encrypted payment vault.
Stack Next.js, Postgres, Stripe Billing and Connect, Discord OAuth, Vercel
Proof 76 production migrations under a production-only release policy
-
Detects retail queue and restock events and fans signed events out to Discord and phones. It detects and alerts only.
Stack Node 22, Python worker, EC2, systemd, heartbeats, APNs push
Proof evidence-based “2 of last 5” rule, replay harness over recorded sessions, a live drop caught and browser-confirmed within 90 seconds
-
A local-first Mac app for profiles, accounts and payment data, with compatibility adapters for third-party formats.
Stack Electron, React, TypeScript, SQLite, Argon2, authenticated encryption, electron-updater on R2
Proof signed, notarized and stapled for Apple Silicon and Intel
-
Records every request, response body, cookie change, WebSocket frame and challenge of a hand-driven browser session into one ordered timeline.
Stack Electron, Chrome DevTools Protocol, Playwright, HAR export
Proof two capture engines share one event schema; exports are redacted by default
Earlier
-
Secure CI/CD and hardened EKS for Department of Defense tenants across development, staging and production.
Stack Terraform, Big Bang, Iron Bank, FluxCD, Helm, GitLab CI, Keycloak, Fortify, SonarQube, Grype, TruffleHog
Proof Big Bang upgrades across every environment; GitOps-auditable deploys
-
Owned the Kubernetes lifecycle of a production Stable Diffusion service.
Stack AWS EKS with GPU nodes, Terraform, ArgoCD, Helm, Prometheus, Grafana, Loki
Proof GPU utilisation and health observable end to end
-
A Go service that probed blocklisted build servers through health APIs and returned them to the fleet automatically.
Stack Go, internal cluster management, a custom CLI for live production build jobs
Proof less manual intervention on Google’s distributed build system
-
A Java tool that flagged anomalous device-registration patterns on Firebase Cloud Messaging.
Stack Java, Flume-based pipelines with PII obfuscation and scoped access
Proof found abusive and misbehaving apps across billions of devices
How I ship
First, freeze the evidence.
Inputs are hashed and immutable, so every run can be replayed against the same bytes.
Then gate on checks, not on hope.
Machine, product and human acceptance are graded apart. A failed check blocks the release.
Give agents a budget.
LLM workers run bounded jobs with retries, a hard spending cap and an honest stop.
Ship it signed, then watch it.
Signed builds, production-only deploys, heartbeats and an alert on every monitor.
Experience
Now
- REDACTED DevSecOps Engineer, Top Secret clearance. Secure systems on RHEL, Red Hat Satellite 6 and Kubernetes.Present
- Independent Founder and engineer, AI products: spatial reconstruction and a production platform with paying customers.2026–
Before
- Raft DevSecOps Engineer, multi-tenant platform for DoD IL2/IL42024–25
- GenAI Infrastructure Engineer, GPU EKS for a production AI service2023–24
- Google Site Reliability Engineer, Build2022–23
- Google Site Reliability Engineer, Firebase (rotation)2021
- Google Corporate Operations Engineer, Techstop2020–22
- Comcast Data Engineer intern2019
- UCCS System Administrator, then Desktop Support2018–19
Education
- University of Colorado Colorado Springs Bachelor of Innovation, Computer Engineering2019
Contact
- Emailben@r4ai.net
- LinkedInbenjamintvigil
- Based inColorado Springs, Colorado