Resume

Applied AI · Decision intelligence · Workflow automation

OpenRate — flagship

Media pricing research, from two working days to about five minutes.

Problem

A national marketing agency's media team priced broadcast radio and TV buys by hand. Their own time study put it near ten working hours per client account, almost all of it lookups — across three incompatible market vocabularies, against a licensed rate card that constrained where the data could live from the first commit.

System

An ad-pricing platform that runs the same workflow as gated steps — ZIP intake, waste, pricing — in about five minutes. It adds a keyless interactive market map, a natural-language query path where a model parses the question into a query spec that a deterministic executor validates and runs (the model never gets to invent a number), editable AI narrative drafts with human sign-off, a budget optimizer, achieved-CPP benchmarks built from 585 real buys, role-gated access enforced at the edge, and a self-serve Data Manager so the team refreshes the rate card without me in the loop.

Role

Sole engineer. Stakeholder discovery with the media team, architecture, implementation, security review, deployment, and the usage instrumentation that measured the result.

Result

• 426 working hours and $27,717 saved across 43 client accounts in the single month of June 2026 — measured by the platform's own usage dashboard, landing within about four hours of the media team's independent estimate
• 225 automated tests plus Playwright smoke specs, a role-permission matrix and 18 screenshot baselines, gated in CI on every push
• A 40-finding multi-agent audit (four read-only agents on separate briefs: security, correctness, data, code quality) worked down to zero open items — including one critical finding refuted with evidence rather than 'fixed'
• Priceable TV markets went 90 → 154 after rebuilding a broken data dimension; 21 crosswalk aliases restored pricing for 16 markets that had been silently unpriceable app-wide

Stack

• Next.js 16 (App Router)
• React 19
• TypeScript
• Tailwind v4
• Vercel
• Vercel Blob
• MapLibre GL
• Claude API
• Playwright
• GitHub Actions CI
• Node test runner
• Python (data ingest)

OpenRate case study

Role-gated — demo on request.

AI-native SaaS · Logistics intelligence · Event monitoring

Calibrate Freight

Turns shipment exceptions into evidence-backed, actionable intelligence.

Problem

A lighting manufacturer's container status came from carrier portals, forwarder emails and a hand-kept spreadsheet. None of them agreed, and none said how it knew. The specific failure mode: a box transships onto a different ship, the booking still names the original vessel, and every downstream ETA is fiction until the container fails to arrive.

System

A multi-tenant container-tracking SaaS where every displayed fact carries a provenance class — OBSERVED, DECLARED, INFERRED, USER_PROVIDED, MATCHED, UNKNOWN — and, where computed, an arithmetic confidence that is allowed to go down. A watchkeeper agent runs on a 60-second tick: it auto-resolves alerts whose condition has cleared, tops up dark vessels with paid satellite fixes under a hard credit cap, and drafts ETA-slip notifications it never sends on its own. Every automated action writes what it did, why, at what confidence and what it cost into an org-scoped ledger that renders in the UI.

Role

Sole engineer. Architecture, the evidence and confidence model, the agent layer, multi-tenancy and isolation testing, deployment and operations.

Result

• Two containers caught switching vessels mid-voyage and auto re-pointed to the correct hull at 0.95 confidence with a plain-English receipt — at zero API credits, because the carrier already knew and the question had been wrong
• 305 web tests and 191 worker tests, CI-gated on a branch-protected main that requires both suites green before a deploy can happen
• Tenant isolation proven with live SQL: an ephemeral second-org user read zero rows across every table, and every write, acknowledge and self-invite escalation attempt was denied
• A 50-agent adversarial UI audit found a headline action that had never once worked for a logged-in user — a permissions gap masked because every real record to date had been created by service-role scripts

Stack

• Next.js 16
• TypeScript
• Supabase / Postgres
• Row-Level Security
• Node 22 worker
• Railway
• Vercel
• MapLibre GL
• AIS streaming
• Carrier container API
• Claude API
• Vitest
• GitHub Actions CI

Calibrate Freight case study

Public live demo at app.freight.usecalibrate.io/demo — client data is isolated in its own org and does not appear.

Applied AI · Sales automation · Decision support

RightSizing Hub

Discovery notes in. Structured, exact-priced plans out.

Problem

Turning HVAC discovery notes into priced, defensible options was manual work.

System

A web application that takes discovery notes and returns three exact-priced tiers — the business's own sizing and pricing methodology encoded as production software behind a tRPC API.

Role

Sole engineer.

Result

• In production and client-owned; test suite green

Stack

• TypeScript
• tRPC
• Web application

Client-owned — described, not linked. Demo on request.