A full-stack app for a grant-writing and government-contracting client, built before orchestrator agents went mainstream. An orchestrator routes each request to the right specialist, and every agent shares one project memory.
Client details private
request
orchestrator
specialist agent
shared memory
Placeholder thumbnail · screenshot pending client OK
PProblem
Grant and contracting work is mostly reading: eligibility rules, long RFPs, compliance checklists. A model can draft it, if it keeps each client's context straight.
DDecisions
Orchestrator plus specialists over one do-everything agent: grants, contracting, business assistant, personal.
One shared project memory so every agent works from the same facts.
A data wall keeps the personal agent (calendar, email) out of business data.
OShipped
Express API, React front end, Supabase auth and memory.
Knowledge base seeded with the client's own data, plus PDF upload.
Drive, Gmail, and Calendar connected through MCP.
Node/Express · React + Vite · Claude API · Supabase · MCP · Railway + Vercel
03 MSB 341 product · BYU Marriott
My SI
A personal model benchmark: it interviews you, turns your real tasks into tests, and tells you which AI model is best for you.
A free, open-source scheduler, designed and shipped for a high-end wedding cake bakery. Clients tap a pickup time and it lands on the owner's Google Calendar, never double-booked.
In progress
github.com/tyleramberry/opencalPlaceholder thumbnail · add a booking-flow screenshot
PProblem
Small businesses pay monthly for a booking link that does one thing. The bakery needed one pickup link and a weekly schedule, without a subscription.
DDecisions
You host it. One link per business, self-hosted, MIT licensed.
Real calendar, race-safe. Calendar busy times block slots live, and a Postgres lock stops two people taking the last slot.
Phone required. So the owner can call about an order.
OShipped
30-minute slots, "Add to calendar" confirmations, and self-serve cancel links.
Confirmation and cancellation emails with the invite attached.
Admin page for hours, address, and pickups.
Deploys to Cloudflare Workers or runs locally on PGlite.
Next.js · TypeScript · Google Calendar API · Postgres (PGlite / Supabase) · Resend · Cloudflare Workers · MIT
A personal operating system I own: a markdown library of what I know, agent skills, scheduled jobs, and an agent team that ships software.
Daily use
350
linked knowledge pages
128
reusable agent skills
7
scheduled jobs
5
agents in the software factory
PProblem
Every new chat starts from zero. I kept re-explaining my classes, clients, and finances, and repeating the same tasks by hand.
DDecisions
Context lives on disk in plain markdown, not one vendor's memory.
Do it once by hand, then make it a skill. Briefs, meeting prep, and inbox triage started that way.
Models judge, code calculates. Scripts do the math, never the model.
People decide, agents do. Shipping, calendar writes, and spending wait for me.
OShipped
Hermes software factory: built on Hermes, the open-source agent platform I ship with. A project manager scopes, product specs, an engineer builds with Claude Code, QA gates every change, and a submission agent ships.
Household CFO agent: bank exports in, budget buckets, weekly check-ins.
Runs on Hermes, Claude Code, or Codex, with an always-on Mac mini for long jobs.