Live · Agent orchestration

Problem: Powerful agent workflows existed but lived in a terminal. Nobody but me could run them, and nothing stopped a bad run from emailing a client.
Built: A local web app over an agent CLI: 21 one-click business workflows, headless streaming-JSON runs, full per-run transcript persistence, and a config-driven registry where adding a workflow is one JSON block.
Guardrail: Draft-only enforced structurally — per-workflow tool allowlists plus a global denylist blocking every send, publish, and delete path. I tested every one before release rather than trusting the prompt.
Result: Operations work that used to need me at a keyboard now runs as a button, and the blast radius of a bad run is bounded by design, not by hope.

Node Agent orchestration Tool-permission design Observability

View source →

Live · Integration layer

Problem: An agent is only as useful as what it can reach. Off-the-shelf connectors didn't cover the systems my business actually runs on.
Built: Three Model Context Protocol servers on a shared modular architecture — a CRM server exposing ~114 tools across 24 modules, a full read/write server over a 57 TB team storage account, and a social publishing server across two platforms.
Hard part: The team-storage namespace made client folders invisible to the API until I found the path-root header; large uploads needed chunked sessions; billing endpoints had undocumented payload requirements I reversed out of failures.
Result: Invoicing, pipeline, scheduling, media, and social publishing all became agent-addressable. This is the layer everything else is built on.

TypeScript MCP REST / OAuth2 Integration architecture

GoHighLevel → Dropbox →

Live · Autonomy

Problem: The expensive failures in a service business are quiet ones — a lead that sat unanswered, a deliverable that aged out, an invoice nobody chased.
Built: A 7:10 AM briefing agent synthesizing two calendars, two inboxes, chat, CRM conversations, pipeline, and outstanding deliverables into one read. A twice-daily lead-response agent that drafts replies in my voice and routes them for approval. A thrice-weekly agent that reviews and improves the skill library itself.
Judgment: The lead agent shipped draft-only with a written graduation gate: two clean weeks before it may propose autonomous send. It never promotes itself. Autonomy is earned on evidence, not assumed on day one.
Result: Response time stopped depending on whether I was on a shoot.

Multi-source synthesis Human-in-the-loop Scheduling Voice modeling

Discovery

Problem: We were paying for an all-in-one business platform and using perhaps 15–20% of it — without knowing which 80% was worth anything.
Built: A read-only walkthrough of every module, sub-tab, and settings tree, turned into a ranked backlog of the ten highest-impact unused capabilities and a three-phase activation plan.
Useful part: I split every item by what an agent can automate via API vs. what needs a human in the UI once — the difference between "I'll handle it" and "here's your click-path." Then I named the features that were real but the wrong priority.
Result: A sequenced plan someone could actually execute, with the distractions labeled as distractions.

Discovery Prioritization Enablement Capability mapping

Live · CRM automation

Problem: Referrals were the #1 lead source and were being asked for by memory, inconsistently, and never tracked back to revenue.
Built: Three published workflows, an intake form, trigger-link attribution, and a tag taxonomy — triggered on final payment, running thank-you, review request, and referral ask on a timed sequence.
Hard part: Getting it to stop. Conditional logic suppresses repeat asks so each client is asked exactly once ever, and retainer accounts never enroll at all. Along the way I found a pre-existing unfiltered notification rule paging the whole team on every task creation, and scoped it down.
Result: Referral revenue became a measurable pipeline segment instead of an anecdote.

Workflow automation Attribution Conditional logic CRM

Live · Buy vs. build

Problem: Equipment tracking across shoots ran on memory and text messages. The off-the-shelf options started around $100/month and did far more than we needed.
Built: A two-operator PWA with QR asset labeling and a custom domain, on infrastructure that costs $0/month at our scale. Access is enforced at the database layer — row-level security plus a signup trigger that rejects any address outside the approved two and auto-confirms those.
Judgment: Access control lives in the database, not the UI, so a bug in the front end can't become a data breach. I onboarded the second operator directly and let him set his own credentials — I never handled them.
Result: Shipped in days, ~$1.2K/year avoided, and a repeatable pattern for evaluating buy-vs-build on small internal tools.

Next.js Supabase / Postgres RLS Vercel PWA

View source →

Method

Problem: Agent-built software drifts. It produces something plausible that quietly isn't what you asked for, and you find out late.
Built: Three chained agent skills — specify, build, verify — with one-to-one requirement numbering, so build coverage and review findings cross-check against each other and the review only passes when every numbered requirement is met.
Result: Evaluation stopped being a vibe check. When a build was flagged as only controller-verified rather than independently reviewed, that gap was recorded as an open item instead of quietly shipping.

Evaluation Spec design Agent reliability

Analysis

Problem: Three systems — accounting, budgeting, CRM invoicing — disagreed about how much money the business made, so every forecast was quietly wrong.
Built: A full reconciliation across all three, plus a books-verified unit-economics model (33.7% variable cost ratio, fixed overhead, interest) tying revenue targets to quarterly milestones.
Found: An entire ~$16.7K/yr revenue channel that never touched the CRM and was understating every projection. Also a scheduled transaction pipeline across four financial institutions, feeding the analysis automatically.
Result: Business debt down from $40.1K to $21.2K across 23 months, on a plan rather than a hunch.

Node / SQLite Financial modeling Data reconciliation

Systems I've built

How I work