A few things I've built.

HVAC service options packet

Problem

HVAC inquiries often arrive after hours, and the owner prices a repair and a replacement by hand before a reply goes out the next morning.

Outcome

The demo turns a homeowner inquiry, and an optional data-plate photo, into a dispatch sheet, three priced good/better/best options from the shop's price book, and a follow-up draft the owner reviews before anything is sent.

Status

Live demo on invented shops, customers and prices. Not client work.

Price book · dispatch sheet · follow-up draft · human approval

Try the demo →

Roof scope packet

Problem

After a roof inspection, the contractor rebuilds the estimate from photos, notes and the adjuster's report by hand, which is slow and makes missed items easy to overlook.

Outcome

The demo turns roof photos and an adjuster estimate into a scope summary, a good/better/best homeowner proposal, a list of possible missed items to verify, and a follow-up draft the contractor reviews before anything is sent.

Status

Live demo on invented jobs, photos and prices. Not client work.

Scope summary · price book · missed-item check · follow-up draft

Try the demo →

Remodel proposal and change-order package

Problem

Remodel proposals are often written at night from photos and walk-through notes, and mid-job changes start as paper time-and-materials tags that the office then prices and packages by hand.

Outcome

The demo turns room photos and a scope note into a low/likely/high proposal with a scope by trade, and a field tag plus photos into a backup-ready change-order package, each with a follow-up draft the contractor reviews before anything is sent.

Status

Live demo on invented jobs, photos and prices. Not client work.

Ranged proposal · change-order package · price book · follow-up draft

Try the demo →

Freight load packet checker

Problem

At a freight brokerage, billing staff check every rate confirmation, BOL/POD and carrier invoice against each other by hand, work out detention and lumper charges from in/out times and receipts, and then key the invoice and the carrier bill into QuickBooks.

Outcome

The demo lines up the three documents in one match table (references, dates, weight, linehaul, fuel, detention, lumper, layover), marks anything missing or unreadable as "needs a look", exports the supported lines as a QuickBooks import file and drafts an email to the carrier or shipper about what's off. People make every call.

Status

Live demo on invented brokers, carriers, loads and amounts. Not client work.

Document matching · detention and lumper checks · QuickBooks CSV · email draft

Try the demo →

Booking and rebooking desk for salons and dental groups

Problem

At a multi-location salon or dental group, every text, call and web request is read by hand to work out the location, service and right person, open times are found by flipping through the book, and regulars who stop rebooking, overdue recalls and recommended-but-unscheduled treatment rarely get followed up.

Outcome

The demo reads each message for location, service, person and timing, drafts a reply that offers real open slots from a sample calendar, and builds lapsed-guest, recall and unscheduled-treatment lists with a draft for each. Unclear messages, complaints, cost questions and any mention of pain are marked "needs a look", and staff approve every message. The dental mode is non-clinical: scheduling, follow-up and estimate clarity notes only.

Status

Live demo on invented salons, dental offices, people, calendars and prices. Not client work.

Message reading · real open slots · rebook and recall lists · human approval

Try the demo →

Order intake checker for home and gift brands

Problem

At a wholesale home-furnishing or gift brand, purchase orders arrive as PDFs and emails from boutiques, reps and designers. Someone retypes every line, catches old SKUs, last season's prices, case packs and short stock by eye, and writes each confirmation or delay notice by hand.

Outcome

The demo checks each PO against the price list, stock and the account on file (SKU, pack, price level, terms, ship-to, ship and cancel dates), flags mismatches and marks anything unclear as "needs a look", exports the clean lines as an order import CSV and drafts a confirmation or delay notice to the buyer. People make every call.

Status

Live demo on invented brands, retailers, SKUs and prices. Not client work.

PO matching · price and stock checks · order import CSV · confirmation draft

Try the demo →

Buyer request → listing matcher

Problem

Buyer inquiries arrive as free text, and someone has to read each one, separate musts from wishes, search the listings and reply.

Outcome

The demo turns a plain-English buyer request into ranked fictional listings with a why-it-matched breakdown, a lead score and a CRM-ready record, and it applies firm limits such as budget as hard filters in code.

Status

Live demo on invented data. Not client work and not a production real-estate product.

Hybrid search · in-browser embeddings · SQL filters · lead qualification

Try the demo →

AI coding-agent migration pipeline

Problem

Dozens of production API endpoints had to move from a legacy service to Go. By hand, each took two to three days and risked changes in behavior.

What I built

A coding-agent pipeline: read-only tracing of each endpoint, generated Go code, mandatory human review and parity checks, and per-route traffic splitting for a safe rollout.

Result

Across about 40 migrated endpoints, most went from two to three days to under half a day.

AI agents · Python · Go · human review

Spec-to-tasks agent planning harness

Problem

Agents turning product specs into epics and tasks burned tokens on raw documents and lost control on long runs.

What I built

A staged pipeline that first builds a verified, source-cited context, then breaks the work down level by level with propose → human approve → write gates, cost ledgers and resumable runs.

Result

Auditable agent runs that can resume after a failure, with a human approval at every level.

LLM orchestration · cost control · approval gates

Database latency fix under high traffic

Problem

A hot read path regressed to about 120 ms at the 99th percentile. The obvious index didn't help, because one value made up about 99.9% of the rows.

What I built

Confirmed the cause with query plans, replaced the index with a partial index on the rare, hot values, and traced the remaining cost to wide JSON columns and connection hold time.

Result

99th-percentile latency went from about 120 ms to about 50 ms.

PostgreSQL · performance · indexing

MCP server toolkit

Problem

Letting an AI assistant use your data and APIs safely needs guard rails.

What I built

A Python Model Context Protocol server with GitHub tools, read-only SQL, HTTP calls driven by an OpenAPI spec, blocked internal-network requests, tests and CI.

Status

Public portfolio project. Works with desktop AI assistants and code editors that support MCP.

MCP · AI agents · Python

Source on GitHub →