In productionA healthcare MSO Brandon founded and sold
Front desk and revenue cycle run by agents
Problem. A seven-location practice group with a two-person office. Scheduling, intake, insurance checks, claims, denials, and payment posting all landed on the same two people.
What we built. A HIPAA-compliant agent platform for insurance verification, claims and revenue cycle, documentation QA, and routine patient requests. Agents handle the routine; a named person handles every exception and every dollar.
Outcome. About 80% of scheduling, intake, and routine support requests resolved by agents. Claim denials fell from about 12% to 1.5%, about $400K a year recovered. AI carried about 85% of finance, operations, and compliance work.
Locally hosted models on the practice's own hardware, FastAPI, Postgres; commercial clearinghouse and billing kept in place
In productionA multi-site education operator
Outbound calls that schedule and capture answers
Problem. Inspection scheduling meant a project manager on the phone with jurisdictions, asking the same handful of questions, then typing the answers into a tracker.
What we built. A voice agent that places the outbound call, asks about project type, jurisdiction, urgency, and contact, and saves the answers as structured data the team can act on.
Outcome. Deployed. Places outbound calls and records structured answers. Outbound only today; the same pattern covers appointment confirmations, no-show recovery, and status callbacks.
Vapi voice agent; answers saved as structured data
Built, dormantA multi-site education operator
Site screening from imagery in a day, not weeks
Problem. Every candidate property meant a consultant visit, a week or more of waiting, and a report that arrived after the good sites were gone.
What we built. Six specialist agents that read drone, aerial, street, and satellite imagery for a site, score it A to F against expert-calibrated criteria, and write specific recommendations.
Outcome. Assessment time went from weeks to same-day.
Gemini vision, LangGraph, Cloud Run
In productionA multi-site education operator
Construction pricing where the AI never does arithmetic
Problem. Early cost and capacity estimates for buildouts depended on who you asked and when. Two people, two numbers.
What we built. A chat advisor that reads floor plans, blueprints, and photos, classifies rooms, checks code and permit timelines, and hands every calculation to a deterministic engine over RS Means data, 37 CSI divisions, and 198 city cost multipliers. Estimates export to a shared document.
Outcome. Same inputs produce the same numbers every time. The AI owns judgment, never arithmetic. Used by the due-diligence pipeline and by other internal tools.
GPT and Gemini, Hono on Cloud Run, Firestore
Built, dormantA multi-site education operator
Construction progress read from scans and video
Problem. Project managers covering many sites could not walk every job every week.
What we built. A system that turns building scans, walkthrough video, and floor plans into a room-by-room report for each trade: percent complete, stage, blockers, punch list, and code checks.
Outcome. Progress and defect reports for PMs without a site visit.
Gemini on Vertex AI, FastAPI on Cloud Run, Firestore
In productionA multi-site education operator
Self-service answers about every building, in chat
Problem. Staff asked the same questions about scan links, square footage, room lists, and files, and one person answered them all.
What we built. A chat bot over the scan library. Plain-English questions return links, files, square footage, and room lists. It also counts fixtures from panoramas, generates interior renders and marketing packs, and places product orders with tracking.
Outcome. Routine questions answered without a person in the loop.
Gemini vision, Google Chat, Cloud Run
In productionInternal, used daily
Follow-ups and commitments that do not slip
Problem. Commitments made in email, chat, and meetings had no owner once the conversation ended.
What we built. An operational assistant that reads mail, files, calendar, chat, and meeting notes, tracks commitments, drafts replies for approval, and runs a morning digest and hourly urgent sweep. Sending, signing, and spending stay with a person.
Outcome. Running on a schedule every day. The same follow-up pattern applies to customers who have not called back.
Claude Agent SDK, Cloud Run, Firestore, BigQuery
Built, dormantA multi-site education operator
Permitting reads the same day, jurisdiction by jurisdiction
Problem. Zoning and code questions for a new site took weeks of consultant time before anyone knew if the site was viable.
What we built. Agents for zoning and code lookup, application drafting, document assembly, plan-check response drafting, and approval tracking.
Outcome. Same-day jurisdictional reads on sites that used to take weeks.
Document-intelligence agents over jurisdiction sources
Aileron