Build log

What I’ve actually shipped

No invented case studies, no borrowed logos, no client names. These are systems I built, run, and pay for, each tagged to the pillar it proves.

Running Scalable Infrastructure

Cast Sense — source-grounded coaching product

A coaching product that answers from a curated source library and live environmental conditions instead of from a model’s memory. Sources are embedded and retrieved with pgvector; answers cite what they came from; and when the sources are thin the product says so instead of inventing something that sounds right.

This is the reference build for the Scalable Infrastructure pillar. Everything I would put in a client RAG pipeline is running here first, on my own budget.

Next.js
Vercel
Supabase
pgvector
Embeddings
Langfuse
Live conditions APIs

Running Secure Acceleration

Guardrails and tracing on a live AI product

Refusal behavior, source attribution, and end-to-end request tracing on Cast Sense. Every model call is traced through Langfuse against an evaluation set, so a confidently wrong answer is something I can find, reproduce, and fix — rather than something a customer finds first.

The transferable piece is the pattern: an AI feature you cannot observe is an AI feature you cannot defend.

Prompt hardening
Source attribution
Refusal behavior
Langfuse traces
Eval set

Prototype Secure Acceleration

Multi-tenant compliance platform

A working prototype for organizations that need tenant isolation and access controls to be verifiable, not just claimed. Core features run and are tested; it is not yet built out to a full production release.

Database-level tenant isolation, encryption in transit and at rest, mandatory multi-factor authentication with role-based views, and immutable read, write, export, and destruction logging.

Multi-tenant isolation
AES-256 / TLS
MFA + RBAC
Immutable audit log

Running Brand Traction

The National Angler content flywheel

An automated content engine: intake collects source material, an AI pass cleans and structures it, a dedupe step kills repeats, and drafts land in the CMS staged for review. Nothing publishes without a human approving it.

The point of the build was to prove that generate-then-review is a workable operating model at volume. The human gate is the feature that makes it safe to run every week. Flo Coast Fishing runs on the same flywheel.

Automated intake
n8n / Make
AI cleanup pass
Dedupe guard
WordPress drafts
Human approval gate

Running Controlled Momentum

Custom dashboards, built and shipped fast

When a client needs visibility into a process instead of a spreadsheet, I build the dashboard, wire it to the real data, and get it in front of them — usually in days, not sprints.

The transferable piece: fast iteration on real data beats a polished mockup of imaginary data every time.

Rapid build
Live data
Fast iteration

Running Controlled Momentum

This site, and the delivery process behind it

A statically generated marketing site with one page per pillar and this build log as the primary asset, because a log of shipped work is a stronger argument than a services page.

Built the way I build client work: written requirements first, agentic execution, human verification, live testing at desktop and mobile before anything went out.

Static build
One page per pillar
Build log first
Written acceptance criteria

Client and facility work is deliberately absent. Regulated-facility engagements involve sensitive security information and are never named or described here, regardless of how good a case study they would make.