A working multi-agent AI operation, built for a real business.
Named agents that hand work to each other through a database. AI shipped inside the products. And the engineering discipline to prove — with numbers — what works and what doesn't.
Pipeline · job-inquiry · demonstrated live
01 · Scout
Finds a wellness-industry job posting via web search, twice daily
02 · Scout
Scores the fit against Amy's profile — only 8+ advances
03 · Sage
Drafts an inquiry email in Amy's voice
04 · Gate
One email to Amy: job summary plus the draft
05 · Amy
Approve · Edit & approve · Revise · Reject — at the Agent Inbox
06 · Memory
Her verdict is written back for both agents to learn from
Three pillars
What this practice is made of
4 agents, 0 dispatchers
routing lives in Postgres
Agent operations
Scout, Sage, Draper, and Squid coordinate through a shared task queue. A single database trigger does all the routing — no orchestrator to babysit, no framework to outgrow.
3 shipped assistants
grounded in live user data
AI in the product
A trading coach that reads chart screenshots. A budget advisor with semantic memory and no server database. An admin assistant with read-only tools. Features, not demos.
−11.3% / trade
the edge we falsified before trading it
Honest engineering
Backtests that report negative results. Bug ledgers that invalidate old stats. Provenance notes on what's original and what's community work. Receipts included.
The work
Three project areas, one operating system
Mindfulness business
The Amy Kuschel ecosystem
A marketing site, a meditation app suite, and a network of AI agents that research, draft, and escalate exactly one decision to a human.
4
named agents on one free-tier VPS
Trading
Journal, coach & backtest lab
An embedded AI trading coach grounded in live trade data, and a research pipeline rigorous enough to kill its own best idea.
5,271
entries backtested against real tick data
Personal finance
Debt Free budget app
A single-file budget calendar with an AI advisor that remembers past conversations — RAG with no server database at all.
287 KB
one HTML file, real-time partner sync
How it connects
One memory, many agents, one human gate
Every agent, script, and coding session reads from and writes to a single governed memory. Content publishes flow in through webhooks, agents recall them before they work, and human verdicts flow back so the whole network learns from one decision.
See the full architecture →Method
Everything starts as a written brief
Every project has a CLAUDE.md, an agent brief, and a plan document before code gets written. Bugs go in a ledger. Claims come with numbers. How I work →
What this means for you
Want a system like this behind your business?
I design and build AI operations end to end: agents, memory, human approval gates, and the product features on top. Every engagement starts with a written brief.
Email me