AI agents and production systems for businesses that run on real workflows.

Planvue

Lead engineer on a multi-tenant platform where venue operators catalog every signage and branding placement on an interactive map: drawn areas, per-location print specs (dimensions, bleed, materials, designer notes), reference photos, downloadable design templates, favorites, and per-venue user management.

DSO operations platform

Principal engineer (contract) on the internal platform a mid-sized dental DSO runs its offices on. The biggest piece is insurance-verification automation: eligibility requests are generated from the schedule, routed by payer to an EDI 270/271 clearinghouse or a headless-browser scraper for that payer's portal, cross-checked, classified (payer answer vs. our failure), and written back into OpenDental, with plain-language failure reasons for the front desk and a Datadog dashboard for the RCM team. The scrapers run through a centralized egress gateway I designed so each office's traffic leaves from its own clinic network, with per-location credentials resolved from 1Password, MFA/OTP handling, circuit breakers, failure capture for post-mortems, and an LLM judge that grades ambiguous outcomes.

Around that: a merge train that batches the volume of AI-accelerated pull requests into hourly runs instead of running the pipeline on every merge, saving about $10,000 a month in CI/CD costs; migration of Rails encrypted credentials to environment-sourced secrets; DB-backed feature flags with percentage rollouts; and an OAuth 2.0 + PKCE layer so Claude and other LLM clients can reach the platform's MCP tools.

Dental Nexus

A bespoke platform for the 20-plus-year pediatric dental practice I help manage, built and run by me as its sole engineer. Because I work inside the practice, the loop is short: a problem spotted at the front desk on Monday can have an automated fix rolled out by Tuesday morning.

It started as insurance verification: browser automation built to run unattended signs in to insurer portals, handles one-time passcodes, matches the patient, and brings eligibility, estimates, and claims back alongside the day's schedule. It grew from there: live call transcription with AI coaching and post-call summaries; an assistant that answers staff questions from appointments and records; a task board fed by email and fax; online booking against the live Eaglesoft schedule (below); daily deposits pushed to the accounting system, with bank transactions matched to insurance payments; and replies to Google reviews drafted by AI and approved by a person.

Online booking for Eaglesoft

Eaglesoft has no API that can write to the schedule, and the office network accepts no inbound connections, so a conventional booking plugin was never an option. I built one anyway: families book hygiene, new-patient, and emergency visits on the practice website against the live schedule, and the front desk confirms from a kanban card. From the first PRD to public launch took five days.

Nexus computes the real openings inside the office, from the Eaglesoft chair schedule minus every block, appointment, holiday, and in-flight online pick, and pushes a PHI-free availability snapshot out to an edge service on AWS (a TypeScript Lambda behind API Gateway, with DynamoDB and SQS). The office only ever connects outward; the edge never reaches in. Holds are written back to Eaglesoft as schedule blocks through my own API, queued through outages and retried with idempotency keys, and watched afterward so a hold that staff overwrite, swap, or lose is reflected on the card instead of silently double-booked.

The rules are the hard part. Every rule is a Who (Eaglesoft appointment types, ages) plus a Where/When (chairs, weekdays, time windows, date ranges) plus one action: don't offer, only offer here, or a limit of n at once, per hour, or per day. Limits count existing Eaglesoft appointments, not just online ones, so two lap exams can never overlap even when the first was booked by phone; children under six are only ever shown mornings, with no hint that afternoons exist; and no booking may leave an orphan gap under thirty minutes. Staff edit rules on a settings page, and rules can only ever remove time.

Practice data and x-ray sync client

Lead engineer on the Windows client Overjet installs in dental practices to feed its AI. It extracts practice-management data on a schedule or on demand, across the major practice-management systems and five database engines, and ships it encrypted to the cloud.

X-rays move in near real time: the client watches the file system for new images, waits for the imaging software to finish writing each one, looks up the patient and exam in the practice's own databases, and uploads the image linked to the right record.