I build production
AI systems end-to-end.

Brad Paws is a pet-care business in SF, and this portal is its system of record. Dozens of clients and their pets run through it, with bookings and invoices processed every month. I treated a small operation as a place to build production patterns properly: I integrated the operator's existing tools instead of replacing them — Notion stayed the workspace, Google Calendar the scheduling source of truth, and the portal took over everything customer-facing. It's a TypeScript monorepo (client portal, API Worker, sync Worker, shared package) on Cloudflare Workers with D1, KV, R2, and Workflows, no origin server. The hard part was making external APIs feel local. I cache OAuth tokens in KV so 98% of Calendar operations never touch Google's auth endpoint.

Clients talk to the portal in plain language. The agent runs on Claude Haiku 4.5 through Vercel AI SDK v6, with fifteen tools wired to the live system: schedule, reschedule, cancel, pull history, send a receipt. Most of the work went into the safety envelope, not the prompt. Destructive tools require two-phase confirmation. There are per-user rate limits, per-conversation token budgets, hard USD cost caps at the user and global level, and a circuit breaker that takes the agent offline when upstream errors spike. I threat-modeled the agent's surfaces (prompt injection, tool authority, data exfiltration) and shipped a mitigation for each before it touched live billing data.

I exposed the Brad Paws platform to MCP clients so I (and authorized agents) can drive the system from Claude Desktop, Cursor, or anything spec-compliant. The server implements the MCP spec and handles OAuth2 with PKCE for client auth. It's stateless across the Workers V8 fleet: every request stands alone, every token is verified, nothing held in memory between calls. This is the one where I read the spec end to end, twice.

Debrief records your interview calls, transcribes them on-device, and scores them against a rubric, so you walk into round two knowing what to fix. The core idea is dual-stream capture: your mic and the interviewer's audio (via ScreenCaptureKit) record as two independent tracks, so who-said-this is a hardware-level fact instead of an ML diarization guess — nothing joins the call as a visible bot. WhisperKit (CoreML/Metal) transcribes fully on-device; the Claude API then produces per-dimension scores, weakness tags, and action items. Paste a company's actual leveling guide and the model weights it above the built-in rubric for that session; an offline mode runs coaching against a local model instead. Audio flushes to disk in chunks during recording, so a crash mid-interview never loses a session, and a trends view charts recurring weakness tags across companies and rounds. Built solo as a native Swift Package with five independently testable targets, each covered by automated tests — including a real end-to-end WhisperKit integration test.

Pawservation is the productization of the Brad Paws scheduler, built solo end to end. A host site adds one <script> tag; Pawservation injects an auto-resizing iframe whose postMessage channel is validated by both origin and source, so neither the widget nor the host page can hijack the other. Behind it is a full multi-tenant platform — each sitter gets isolated config, services, pricing, and bookings, with capacity and conflict rules computed from a single source of truth, plus a non-technical dashboard for confirming bookings, defining time-windowed services, importing clients from CSV, and syncing to Google Calendar. The booking/date/pricing core is pure TypeScript with zero runtime dependencies, so the rules that decide what a customer can book are testable in isolation — and tested against in-memory SQLite. CI gates every PR on typecheck, lint, format, test, and build, then auto-deploys to Cloudflare on merge.

MCP Auth Kit packages the hard parts of shipping a safe Model Context Protocol server so others don't rebuild them from scratch. It handles OAuth 2.1 with PKCE, rate limiting, scope-gated tool access, and two-phase confirmation on sensitive actions, and stays unopinionated about your tools, identity provider, and storage. It's the reusable, open-source distillation of the auth and safety work behind the Brad Paws MCP server.

Log in with Strava, link an eBird or iNaturalist profile, and Roadrunner writes the species you logged into the description of the matching Strava activity. Activities sync in real time via Strava OAuth and webhooks, with automatic token refresh. The interesting problem: birders submit checklists hours after they get home, so unmatched activities go into a deferred re-check queue (2/4/8-hour ladder), and the species appear the moment the checklist does. Deployed serverless on Vercel with pooled Neon Postgres and scheduled by a bearer-authenticated GitHub Actions cron. Three third-party API integrations, timezone-correct wall-clock matching, security middleware (CSRF, clickjacking, per-user webhook rate limiting), and a test suite covering matching, sync, OAuth, and the re-check queue.