It's client data, private tools and the systems I run my own business on, so it can't go on a public page. Here is the evidence instead: real file names and real output from my repos. Pick any one and I'll walk you through it live.
Grant platform
The Grant Greenhouse
A grant operating system. It reads who holds federal money and when it runs out, works out which foundations have to pay out soon, and builds the funder-ready package. Every claim in the package is tied to its source, so an unsourced claim can't print as fact.
db/migrations
004_rls.sql client isolation
007_payout_pressure.sql payout pressure
008_recompete_signals.sql award expirations
013_packet.sql funder packet
020_grants_990.sql IRS 990-PF data
23 migrations in all
A private portal where each client sees only their own workspace. My own contact form feeds a lead table, emails me with a lead number and sends the sender a confirmation. Same pattern, built to fit the business.
client portal (login required)
one workspace per client
requests board
projects and milestones
files
advisor (t.ROY)
owner can step into any workspace
Multi-tenantNeon PostgresMagic-link login
Agents
A team where each agent has one job
Besides t.ROY, I run eight research agents. One of them exists only to attack the others' claims before anything reaches me.
adversary attacks claims
extractor pages into records
news-watch reports what changed
signal-scout proposes signals
outlier-hunter finds what's off
pattern-weaver connects domains
barbell-assessor park, probe, push
briefer decision-ready briefs
I design the schema, the access rules and the migrations, and pick the right store for the job. Steel Man Resumes runs on 74 migrations in a public repo you can read yourself.
packages/core/migrations
001_foundation.sql
019_client_code_tier.sql
059_owner_only_rls.sql
073_resource_verification_tiers.sql
74 files in all
I ran a pre-registered study on how AI picks who to recommend. I also test my own agents. I asked t.ROY ten real questions, rewrote its rules and asked again. The average answer went from 101 words to 51.
question before after
why trust a consultant... 112 40
is my client data safe... 176 53
are you any good... 113 35
small HVAC company... 94 33
10 questions, average 101 51
A build guard that blocks bad pages, a shared header and footer stamped onto every page, a checker for writing that sounds machine-made. I write them in Python, Node, SQL and shell.
$ node scripts/preflight.js
preflight: 0 error(s), 0 warning(s)
STATUS: PASS
$ python3 scripts/chrome.py --check
chrome OK on 25 pages
“He is thoughtful, proactive, and consistently makes suggestions that help move my business forward without ever being pushy or making me feel like I'm losing control of my own brand.”
