Selected work
Production systems, the framework underneath them, and the public libraries they feed. Everything on this page is running, public, or verifiable. Client engagements are covered by confidentiality and are described by shape, never by name.
From the GitHub ledger published with the CV, generated 2026-09-03, covering 14 repositories from December 2025 onward. One engineer, working solo, with agentic tooling.
Built and operated end to end: architecture, implementation, CI, deployment, and the on-call that follows.
Architecture · implementation · operations
A public AI terminal that is neither an open chatbot nor decoration: structured intake routes each visitor into a persona mode before the first model call, and the backend owns every decision after that. It runs as a live prompt session, bilingual in English and es-CO Spanish.
Control plane · doctrine · 623 commits since 2026-06-11
The factory the rest of this page is built in. A control plane for running a one-person software studio at multi-team throughput: a native operator cockpit, a token-gated web panel behind it, and a worker fleet that keeps the queue moving while the operator is asleep.
Framework extraction · public repository
The terminal generalized into a reusable, brandable framework: FastAPI backend, Next.js terminal, and deployment bundles that make branding a directory of configuration instead of a fork. Extracted from the production system in May 2026 and developed in the open since.
Architecture · implementation · 1,007 commits since 2026-05-28
A self-generating smart LMS. Name a subject or upload your own documents; the system researches it, bounds it in cited sources, designs a curriculum grounded in learning science, generates the materials and assessments, grades its own output, and publishes a branded learning portal that runs on your infrastructure. A learning system, not a quiz site.
Curated public library · automated pipeline
A Git-backed catalog of community-built automation and GenAI-Ops content, assembled by a scheduled discovery pipeline and curated by explicit operator review. Human-browseable on the site and machine-readable for AI crawlers.
PowerShell library · observability and testing
A modular PowerShell automation library for MSP administration, built with the operational furniture such scripts usually skip: structured observability, security boundaries, and a test suite.
Assistant skill · public repository
The study loop used to earn the Anthropic certifications this page claims, released as a skill: profile, baseline diagnostic, a phased plan sized to a real calendar, and a persistent miss log. Three lanes under one coach: Architect Foundations, Architect Professional, Developer Foundations.
Assistant skill + local dashboard · public repository
A job and contract search run as a loop rather than a panic: guided intake, honest 0–100 scoring, live-link verification, and a five-lens adversarial resume review per role. No server, no signup, and it never auto-applies.
Full-stack build · real-time multiplayer
A real-time multiplayer AI party word game: synchronized rooms, live scoring, and model-generated play running against latency budgets a turn-based system never has to meet.
Program Module on GWEDU · bilingual (EN/ES)
An AI lifestyle and fitness coach shipped as a Program Module on GWEDU: its own repository of content and configuration, zero engine code. It is the proof the module model works: a completely different subject domain, brand, and persona set running on the same engine as a training course, with no fork.
The parts a demo does not reveal. Each of these is a decision that survived contact with production.
Public AI features fail on cost before they fail on quality. Every cloud call passes a budget gate with call caps and hard dollar ceilings, and visitors choose the question, never the provider or the spend.
The stronger guarantee is the ledger: a completion that cannot report readable token usage trips a persistent breaker that takes cloud routing offline until an operator clears it. That breaker has fired in production against a perfectly healthy response with a changed usage format, which is the design working. Unmetered spend is an outage, not a rounding error.
The public status indicator is computed from whether a route can actually serve the next turn, not from whether keys happen to be configured. A system that says "online" while every route is dead is worse than one that admits it is degraded.
Retrieval quality is where hand-tuning goes to hide. The search behind the terminal is measured against a 73-question golden set built from real questions, each paired with the specific production chunk that should answer it, run over the live 63,105-chunk corpus. Adding a cross-encoder reranker moved recall@3 from 0.42 to 0.51, and that is the number that justified shipping it.
The same harness deleted work. Two hand-written query classifiers that nudged scores by substring looked reasonable and measured as noise: one fired on 0 of 73 questions, the other on 7, six of them false positives. They were removed rather than defended. The golden set and the eval script now live in the repository, next to the code they measure, and the harness fails loudly if it ever drifts from the search path it claims to be scoring.
Palette, logo, identity, intake questions, personas, routing policy, and copy load at runtime from a deployment bundle. One codebase serves a consulting terminal and a fitness coach without a branch between them, which is what makes the framework worth releasing publicly at all.
A scheduled discovery pipeline polls feeds, repositories, topics, forums, and video channels; a scout scores what it finds; notable items flow to a public ticker, a curated digest, and the CODEX library. Featuring and publishing stay explicit operator actions with audit records. The automation gathers, a human decides.
Engineering practice, stated plainly, because it is the part that survives the handoff.
4,648 backend tests run on every pull request alongside frontend lint and build and an automated code review gate. A merge to master deploys production on its own, but only once CI on that exact commit is green; the manual path still exists for reruns and emergencies and still requires a typed confirmation.
Model access runs through an on-host key broker with short-lived virtual keys, so rotating a provider credential is an operations task rather than a code deploy. Credentials never enter application code, logs, or model context.
Scheduled jobs, backups, restore verification, and retention run on the box with their own logs and alerting paths; admin actions are token-gated and audited. The unglamorous half of shipping is treated as part of the build.
Verifiable, and linked to their issuers.
CCAR-P, issued by Anthropic, 2026-07-17.
CCAR-F, issued by Anthropic, 2026-07-01.
CCDV-F, issued by Anthropic, 2026-08-02.
All four are on one Credly profile, and the full history is on the CV.
Client engagements are covered by confidentiality. Detail on scope, structure, and references is available on request under NDA.