I spent years running business operations — finance, CRM, multi-entity workflows — and kept hitting the same wall: critical processes held together by manual work and spreadsheets. So I learned to build. Now, when I see an operational bottleneck, I design and ship the system that removes it: from workflow mapping through architecture, code, deployment, and adoption.
I'm not a classical algorithms engineer. I'm the person who understands the business problem, opens a laptop, and builds the end-to-end solution — with production discipline: typed code, tests, CI, human-review controls, and safe failure modes.
The problem: Leadership at a multi-entity AdTech company relied on four disconnected systems — a performance API, Monday.com CRM, Google Sheets finance ledgers, and an operational database. No single view showed what was happening, whether the data was fresh, or where to intervene.
The build: A unified executive command center: typed connector contracts isolate each provider's schema, a snapshot orchestrator pulls all sources concurrently with independent health reporting, and a business-rules layer turns raw events into KPIs, alerts, and goal pacing. In the production system this runs on scheduled syncs with self-healing backfills — and serving optimization cut dashboard TTFB from 3.4s to ~30ms.
What it demonstrates: Multi-system integration architecture (CRM, finance, performance data), operational reliability engineering, and translating executive decision-making needs into a data platform. The public repo is a clean-room edition with synthetic data.
Stack: Next.js · TypeScript · React · Monday.com API · Google Sheets API · Supabase · Vercel Cron · Vitest · GitHub Actions
The problem: Music teachers see students once a week, but the learning happens between lessons — and that practice is invisible. Manual practice logs add friction and rarely reflect what was actually played.
The build: An end-to-end B2B2C system. A Raspberry Pi captures MIDI from the piano, a FastAPI backend compares each performance against the uploaded sheet music (MusicXML parsed with music21), and GPT-4o turns the structured error report into two tailored outputs — a technical report for the teacher, an encouraging one for the student.
What it demonstrates: LLM integration on top of deterministic analysis, edge-to-cloud architecture, and role-specific product thinking.
Stack: Python · FastAPI · PostgreSQL · OpenAI GPT-4o · music21 · Next.js · TypeScript · Raspberry Pi
The problem: Finance teams burn hours copying invoice data between systems — and a single bad sync can silently overwrite valid financial records. Built from real multi-entity finance-operations experience.
The build: A reconciliation pipeline that treats sync as a controlled workflow, not a blind import: deterministic diffing by stable IDs, automatic handling of safe changes, and human-review routing for suspicious ones (amounts zeroing out, sign flips, material changes). Dry-run by default; nothing is written without an explicit --apply.
What it demonstrates: Automation with auditability and guardrails — the difference between a script and a system a finance team can trust.
Stack: Python · Pydantic · Google Sheets API · pytest · mypy · GitHub Actions CI
- Start with the operational problem and the people living with it
- Map the workflow, constraints, and failure modes
- Build the smallest system that creates real leverage
- Ship with controls: typing, tests, dry-runs, human review where it matters
- Deploy, measure adoption, iterate
Languages: Python, TypeScript, SQL
AI: OpenAI API (GPT-4o), LLM workflow design, prompt engineering for structured outputs
Backend & data: FastAPI, PostgreSQL, SQLAlchemy, Pydantic, Supabase
Frontend: Next.js, React
Integrations & automation: Google Workspace APIs, Monday.com, REST APIs
Delivery: GitHub Actions, CI/CD, Docker, Alembic, Raspberry Pi
Looking for roles at the intersection of AI, operations, and implementation — where understanding the business is as important as shipping the code.