Proof, not promises.

CASE_01

Marshal: AI candidate sourcing for a recruiting firm

Result: A paid AI sourcing platform, live in production and co-owned with the client, that turns a job description into a ranked, evidence-backed candidate shortlist.

The problem

The firm's head of sales sourced candidates by hand with Boolean searches, one role at a time. It was slow, and the reasons a candidate made the list lived in his head rather than on the record.

What I built

• Job reading: Marshal reads a job description and proposes the requirements that matter, separating required from preferred. The recruiter approves them before any paid search runs.
• Search and scoring: it searches candidate data providers, scores each person against the approved requirements, and ranks the pool, citing the evidence behind each match.
• Guardrails: it never infers sensitive qualifications, such as a security clearance from an employer or job title, and it records whether a fact was "candidate stated" or "recruiter verified."
• Recruiter workflow: results on screen, outcome marks, and Excel/CSV export. Contact details are bought only for people the recruiter selects.

How I verified it

• Acceptance test: before seeing Marshal's rankings, the client built his own shortlist by hand from the same candidate pool, marking each person In, Maybe, or Out with a reason.
• Contract per batch: written acceptance criteria, an independent AI code review, and my sign-off before every release.
• Client walkthroughs: his feedback became a four-step flow: paste the job, check the requirements, press Find people, review and export.

"Marshal does the first pass I used to do by hand with Boolean searches, and it shows me why each person made the list. I approve the requirements; it does the searching."— Head of Sales, recruiting firm

> requirements · acceptance testing · AI evaluation · cost controls · responsible AI · deployment

CASE_02

AI business assistants for two small-business owners

Result: Two managed AI assistants live for a barbershop and an independent insurance agency. Each passed every test before go-live, and each drafts work in the owner's voice but never sends anything without approval.

The problem

Both owners run their businesses themselves. The barber had no time for social captions, review replies, and customer messages. The insurance agent spent hours writing lead follow-ups and deciding whom to call, in a field where one wrong sentence about price or returns is a compliance problem.

What I built

• Barbershop assistant: drafts Instagram captions with hashtags and posting times, drafts replies to Google reviews and customer messages, answers questions about the shop, researches local competitors, and gives weekly summaries.
• Insurance assistant: builds a prioritized daily call list, drafts lead follow-up texts and post-call email sequences in the owner's voice, works through common objections, and drafts Facebook ads that follow the platform's rules for insurance advertising.
• Hard limits: both draft only and never send or post on their own. The insurance assistant never quotes prices or gives investment advice, and flags leads from states the owner isn't licensed in.

How I verified it

• Wrote the specification and test rubric before building, including a 15-section specification for the insurance agency.
• Ran boundary tests designed to push each assistant to break its rules. Asked for a monthly price to text a prospect, the insurance assistant declined and drafted a message offering a 15-minute call instead.
• Audited both assistants' permissions so any tool able to send, pay, delete, or publish must ask first.

"I'm behind the chair all day. Now my captions and review replies are ready when I am, and nothing goes out unless I approve it."— Owner, barbershop

"Every morning I know who to call first, and my follow-ups sound like me. It also knows what it's not allowed to say, which matters in my business."— Owner, independent insurance agency

> client onboarding · AI testing · compliance-aware AI · permission design · training non-technical owners

CASE_03

Herald: governed AI publishing

Result: A live social-media publishing platform where AI can draft posts and images, but nothing reaches the public without a named human's approval, every release is logged, and AI spend has a hard cap.

The problem

Businesses want AI help with social content, but AI can post the wrong thing, overspend, or publish images people can't tell are AI-made. Most tools treat approval and audit as afterthoughts.

What I built

• Human approval gate: every post and AI-generated image needs a named approver before release.
• Audit ledger: every release is written to a permanent record showing who approved it and what went out.
• Spend governor: AI image generation runs under a hard daily dollar cap; over the cap, nothing is spent.
• Fail-closed publishing: if a platform's response is unclear, Herald holds the post and checks its status rather than assuming success.
• Disclosure: AI-content disclosure and watermark notes travel with each record.

How I verified it

• Every safeguard has a test proving it blocks the bad case and a test proving normal work still goes through.
• New capability ships switched off; turning on live AI generation was a separate, deliberate step with its own checklist.

Outcome

Herald is live and runs Terminal X's own social media across five platforms. Every post is approved by a person before it goes out, and AI spending stays under its cap.

> AI governance · human-in-the-loop design · cost controls · audit · release management

CASE_04

Brain Bridge: governed memory for AI-assisted work

Result: A local system that gives every AI work session on my PC a dated briefing of where each project stands, and gives cloud sessions an audited digest. Privacy and secret rules are enforced in code, and each one is proven by a test that fails when the rule is switched off.

The problem

My project knowledge lived in a notes vault, about 40 project folders, and cloud AI sessions that can't see my computer. The vault's status file was five months out of date. Worse, the AI coding tool was quietly deleting session history after 30 days, and my main project had six sessions left.

What I built

• Read-only connectors: one reader per source pulls only what's new since the last run and never changes a source file.
• Dated search: every note, decision and session turn becomes a dated record. Search ranks by text match, with newer records ranked higher.
• Session briefings: each new AI session on my PC starts with a short briefing for its project: the last real status, recent decisions, and notes waiting to be filed. It is framed as background facts, never instructions.
• Cloud digest: a small audited copy for cloud sessions, pushed to one private repository.
• Write-back inbox: sessions leave dated notes; I approve each one before it is filed.

How I verified it

• Privacy quarantine: one folder holds other people's personal data. It is never read, and any AI session that mentions it is excluded whole, permanently. I first ruled a looser 3-turn window, then reversed it the same day.
• Audit before publishing: every stored record is re-scanned after every run. If anything secret-shaped or protected appears, nothing is published. A deliberate test with a fake key planted in the build was blocked.
• Switch-off tests: every rule has a test that turns red when the rule is switched off. 83 rules were checked this way.
• Adversarial reviews: after four of the five batches, a fresh-context AI reviewer got only the code and tried to break it. The four reviews found 9 high-severity issues, and all 9 were fixed before release.

Outcome

In its first 39 logged runs: 0 crashes, and every stored-content audit came back clean. The few alerts were health checks doing their job, such as catching a mistyped folder. Briefings load in under 0.2 seconds. The upkeep budget is 0 minutes a day and about 10 minutes a week. A sanitized version of the code, tests and decision log is public on GitHub.

> requirements · AI governance · privacy by design · testing and verification · knowledge management · agent orchestration