I build automation systems for businesses that run on paperwork, plus a handful of products of my own — usually the ones I got tired of waiting for someone else to build. I'm opinionated about the software I use, and most of what's below started as something that bugged me.

Most of it ships private — the contribution graph is real, the source just isn't public.

How I build: deterministic core first — tested extraction, tested math, repeatable output — then non-deterministic workflows layered on top, only where judgment actually helps. The model goes at the edges, never in the critical path.

🌐 Portfolio: dmarcinowski.com  ·  ⚙️ Shop: automagic.ly

Client work · live since Feb 2026

Monthly telecom invoicing for a telecom-heavy client, end to end.

• 33 carrier accounts, 22 retrieved unattended — portal automation, reimplemented HTTP logins, or straight out of an inbox, because no two carriers deliver the same way
• 22 extractor formats normalize the bills and split charges across cost centers
• Cover sheets, monthly reporting, and submission back into the customer's AP intake system — all of it automatic
• Retries know the difference between a failed download and a carrier that just hasn't posted the bill yet — different problems, different escalation
• Operated, not just deployed — CI/CD on push, cron across the billing cycle, a daily failure digest by email, and 11 command-level runbooks written precisely enough that an agent runs the monthly support cycle
• Downloading, coding, cover sheets, review: the old way took about 20 hours a month

Python · FastAPI · SQLite · Playwright · React · Fly.io

Client work · live since late March 2026

QR-badge time clock that replaced paper timesheets at a small clinic — kiosk clock-in, employee management, hours dashboard. Offline-tolerant PWA, so a dropped connection never costs someone their punch.

Live

An online résumé you can cross-examine — and one that's allowed to say no.

• Every résumé has been perfect since 2023. What used to be signal is now noise: a flawless document proves nothing when a flawless document takes ten seconds
• So this one takes questions instead. Paste a job description, interrogate the work history, get a fit verdict that can come back "probably not" — because an assessment that can't say no can't meaningfully say yes

Next.js · Supabase · Claude API — dmarcinowski.com

Solo build · private beta · in daily use since Jul 2026

Training analysis for runners — the dashboard that talks back. The watch records; this is where you sit down and think about it. No kudos, no feed, no post-run dopamine.

• A coach that already knows you — two-way MCP puts your whole training history in the room at the start of every conversation. No screenshots, nothing re-explained, no starting over. Its decisions land straight back on the dashboard
• Local-first on purpose — no accounts, no servers, no subscription. Your training history is a SQLite file on your own machine, synced outbound-only under your own credentials
• Runs on the AI you already pay for, not a second subscription
• Approved on the Suunto API partner program

Electron · SQLite · MCP — debrief.run

Open source · in development

Photo-to-macros isn't new. Owning it instead of renting it — and having it know what I ran this week — is.

• Model at the edge, math at the core — Sonnet 5 vision estimates macros from a photo; the daily budget is deterministic, computed against real training load
• Run data arrives through Debrief's Suunto pipeline, weight from a Garmin Index scale — the two are built to meet
• The one public repo here — everything else on this profile is closed

React · Hono · Cloudflare Workers · D1 · R2 · TypeScript — mymacros.debrief.run  ·  source — clone it and run your own

In development

Shows a high-school distance runner where their times actually stand against college recruiting benchmarks, tracked over a multi-year horizon.

• Benchmark math is deterministic and tested, graded against an expected age curve — so normal development never reads as failure
• Runs on 📚 RunDB, a data engine with per-record provenance: every source passes policy review before a single fetch, and the audit trail is committed alongside the code

Landing page live · v2 in progress

The shop — practical AI-assisted process automation for businesses that run on paperwork. The v2 site ships with a proof-ledger content model: every marketing claim carries its source, method, and caveat before it goes live.

• 🖥️ local-AI — an M4 Pro running Qwen3-Coder on MLX behind an outbound relay, giving a community a 24/7 agent at zero marginal cost per token
• 🗃️ theDBSample — provenance-tracked research database for collegiate track & XC results; ships the pipeline, never the data
• 🧠 Memex — an LLM-maintained knowledge base holding only what a model can't look up: decisions with their reasoning, project context, hard-won gotchas

_{Active work is private for now — ask me about it.}