BUILD LOG
WHAT I’M BUILDING RIGHT NOW, AND WHAT I SHIPPED RECENTLY
NOW
PICKUP
Building · Started Oct 2026
An AI front desk for small businesses. It answers calls, books jobs, and follows up, so owners stop losing customers to voicemail. Owners log into a dashboard to read every call transcript and update their hours, pricing, and services. The agent always says it’s AI if asked, and hands off to the owner when it should.
Where it is: dashboard v1 built on mock data.
Next: connecting a live voice agent.
It grows out of Dispatch, my hackathon project, and Shift Rescue.
CHEF DESK
In testing with one chef · Started Oct 2026
A planning app for private chefs. Clients, weekly menus, and grocery lists live in one place instead of the Notes app. AI suggests dishes based on each client, and the chef approves everything. Built from discovery calls with a working private chef who plans every week by hand.
Where it is: live and in use by its first chef.
Next: one-tap “send menu to client,” so chefs can send a proposed week by text.
Stack: Supabase · Vercel · Fireworks
KILL SIN
Paused since Sept 2026
An iOS accountability app with an AI mentor, Shepard, for Christians fighting one specific habit at a time. It runs on my iPhone as an internal build, with Shepard on DeepSeek via Fireworks.
Next: a free TestFlight beta for a small group.
→ Full build history
LOG
OCT 2026 · SHIFT RESCUE
A voice agent that backfills shifts when a driver calls out. It works down the list of available drivers by phone, confirms the first yes, and fills the slot. I adapted it from Dispatch’s refill engine, using ElevenLabs for voice, and deployed it on Vercel.
Voice AI · Agents · Ops automation
→/shift-rescue
OCT 2026 · STOREFRONT TEARDOWN AGENTS
Four Claude Managed Agents that crawl an online store, audit it with evidence, rebuild it two ways, and package a ready-to-send report. A person approves each stage. On one run, the design agent caught a wrong audit finding, and the report agent dropped it before it reached the final report. About $9 per full run.
Multi-agent · Claude Managed Agents · Next.js
→ /storefront-teardown-agents
SEP 2026 · DISPATCH · HACKATHON FINALIST
A voice agent that refills cancelled appointments by calling a business’s waitlist one person at a time. I led agents and orchestration on a team of four: two Claude Managed Agents, the refill engine, and five end-to-end test scenarios. Live testing caught four voice-agent bugs, and we fixed all of them before I demoed it live on stage. Built in one day at Dream AI Hackathon 2026.
Voice AI · Orchestration · TypeScript
→ /dispatch
SEP 2026 · STORY & SPOON CLIENT SITES
Three sites built through my studio’s config-driven template: a San Mateo dessert lounge, a Bay Area sneaker reseller, and a private chef. Each runs on one config file, so a new client means new content, not new code.
Web · Brand systems · Claude Design
→ /storyandspoon
SEP 2026 · VC GROUND TRUTH
Turned a VC investor’s board decks and KPI spreadsheets into a database he can question in plain English through Claude. It holds 31 documents, 25,000+ metrics, and 1,269 embedded pages, and every numeric spot-check matched exactly. A watcher ingests new files automatically.
ETL · Postgres/pgvector · MCP
→ /vcgroundtruth
AUG 2026 · KILL SIN: FIRST iOS BUILD
First real device build through EAS, after a 29-issue launch audit: server fail-closed, request timeouts, dev tools stripped from release builds, and a 510-test storage suite.
React Native · Expo · Launch hardening
→ /killsin
AUG 2026 · SHEPARD V3 + FIREWORKS MIGRATION
Rebuilt Kill Sin’s AI mentor using evals. I ran a 4-model bake-off and moved production from fine-tuned GPT-3.5 to DeepSeek on Fireworks, which cut cost about 20x. Crisis referral is now enforced in code: 15/15 escalated conversations refer out, with zero false triggers.
Evals · Model migration · AI safety
→ /killsin
AUG 2026 · JARVIS V1
A local-first personal AI system that ran my job search: Gmail read-only, an approval inbox, and 498 tests. Its rule: observe freely, propose everything, send nothing. Retired Oct 2026.
Agents · Human-in-the-loop · Next.js
→ /jarvis
JUN 2026 · CAMPAIGN RISK COPILOT
A finance review tool for ad campaigns. Code does the math, and AI only explains it. Missing data shows “N/A, cannot be computed” instead of a fake $0.
Finance automation · Deterministic + LLM
→ /campaign-risk-copilot