Work
Selected shipped work, client work, and labs. Status matters.
Drew Garman builds practical AI software and websites. Key projects include DBNT, an open-source Python protocol for AI feedback loops (available on PyPI), and Borussia Minerals, a client e-commerce site for premium mineral specimens.
Borussia Minerals
liveClient: Boris Dimov
Client delivery: e-commerce + 3D Gaussian-splat viewer + Stripe
Challenge
A premium mineral specimen dealer needed a professional web presence that matched the quality of Arizona's Fat Jack Mine specimens — including a way to show off 3D point-cloud captures in the browser.
Solution
Designed and shipped a full-stack e-commerce site with a custom 3D Gaussian-splat viewer, Stripe checkout, and responsive museum-quality UI. Built and deployed solo, end-to-end, for a paying client.
Result
Live at borussiaminerals.com. Real paying client, production Stripe, 3D splat viewer running in-browser. Demonstrates client delivery ownership: spec, design, build, deploy, maintain.
Family Recipe Box
liveFull-stack product with real users — auth, multi-tenancy, email, AI transcription
Challenge
Build a multi-family recipe platform that's genuinely private (each family sees only their recipes), handles real-world auth edge cases, and works on mobile.
Solution
Full-stack Next.js app with PostgreSQL, per-family tenancy, Firebase keyless broker + Better-Auth sessions, magic-link email via Resend, and a Mama-node AI transcription worker for scanning handwritten cards.
Result
Real users in production. Caught and fixed a hairy production bug: ISP scanners (cox.net) were pre-fetching the magic-link URL and burning the token before the user clicked. Fixed with a POST-gated confirm interstitial. This is the kind of debugging FDE work is made of.
Abledumb
liveAI-assisted music workstation with a local-first router
Challenge
Most music tools either bury producers in menus or try to replace them entirely. Build a workstation where the artist stays in control and AI helps with the boring or blocked parts.
Solution
AI-assisted media editor with local model routing through Ollama and LM Studio when the task allows it, plus cloud APIs only where frontier capability is actually needed.
Result
Live at abledumb.com. Demonstrates the product principle I keep coming back to: AI should accelerate the producer without taking ownership of the track. CSP was hardened from report-only to enforcing after live-traffic review.
GUS — 4-Node Local AI Fleet
liveLocal-first AI infrastructure for routing, governance, and agent work
Challenge
Running every AI task through a cloud API is expensive, brittle, and wrong for sensitive work. I needed a way to route routine tasks locally and save cloud models for the jobs that need them.
Solution
Designed and operate a 4-node Apple Silicon AI fleet (GUS boss node + Mama compute drone + DJ worker + DRU workstation) with local/cloud routing and a runtime governance layer enforced by hook scripts.
Result
This is the infrastructure layer I use to test local routing, agent handoffs, and governance before turning them into product or client work. It is lab infrastructure, not a finished product claim.
DBNT — Do Better Next Time
liveOpen-source AI feedback protocol — published on PyPI
Challenge
AI agents repeat the same mistakes across sessions. Traditional error logging captures what failed but not why — and success patterns are ignored entirely.
Solution
Open-source Python protocol encoding both success and failure patterns with weighted signals (success 1.5x over failure), FSRS spaced-repetition decay, and pluggable adapters for any AI framework.
Result
84 tests, zero external dependencies, pip-installable. Published on PyPI. Demonstrates systems thinking about AI improvement loops.