Proof of Work

The flagship

Proof Engine: an AI operations platform that runs my business

Most people have AI chatbots. I built an AI workforce. Proof Engine is a platform I designed and wrote from scratch that plans, drafts, schedules, and dispatches the growth work for eleven business projects, with me approving the output instead of producing it.

proofengine.bizinabox.online

Live Node 20 · Express 5 · SQLite · Claude Sonnet 4.5 via OpenRouter 160+ tests · 26 test files

Every morning at 9am an AI Chief Strategist reviews every active project: queue ages, executor throughput, funnel state, live metrics, what content actually performed. It sends me one briefing, proposes at most two moves with evidence, and tags each piece of work by who should do it: the AI, me, or a hire.

It learns from results. It traces posted content back through the proof it was built on, so it can tell me "response-time stories run 1.4x your average." It flags any signal built on fewer than 3 data points as early, not proven.

AI executors do the actual work. A worker process wakes every 10 minutes, picks up queued items assigned to AI staff, and executes them through a skill registry: drafting outreach, triaging inbound replies, generating images, building video production kits. New skills are added to the registry over time. That is how the AI workforce learns, and it is the same pattern I would use to scale a human team.

Nothing leaves without a human. All AI output stops at an approval inbox, and the gate is enforced server-side at the dispatch layer, not in the UI. Autonomy is a per-agent toggle that defaults to off. That architecture decision is the one I would defend hardest in an enterprise.

Voice system that learns from rejection

Reject a draft with "I don't say that" and it extracts a permanent style rule, deduplicates it against existing rules, and enforces it with a deterministic detector on every future draft.

Outreach funnel + LinkedIn automation

Contacts move new → contacted → replied → interested. AI drafts every touch; SendPilot and GoHighLevel dispatch after approval; inbound replies auto-queue an AI triage run.

Conversational Telegram operator

A tool-using AI assistant locked to my chat: search the proof vault, assign work, log results, draft posts in my voice, all from my phone.

YouTube comment triage

AI sorts every comment into leads, questions, praise, criticism, spam; drafts replies in my voice; escalates what needs me personally at 7am.

Vision-grounded strategy

Product screenshots are passed to the model as images so the strategist reasons from the real product, not a description of it.

Agent-drivable by design

A CLI and a one-call briefing API expose the whole operation, so an AI coding agent can run the business from a terminal.

The portfolio

Production AI systems

Each of these is a working system, not a demo. Status is stated honestly: live, built, parked by choice, or killed. Every live link below returned HTTP 200 when checked on August 25, 2026.

ParseFlow

Built, pre-launch

Multi-tenant AI invoice parsing for a barter trade network whose invoices split every amount between cash and trade credit. Documents arrive by forwarded email, upload, or API; up to 10 parse concurrently into a human review queue.

AI roleClaude Opus reads invoice page images by vision and returns structured data. Quality control is zero-token arithmetic validation plus a random 10% human spot check.

First tenant: MK Global Trade (live client)

StoryHero

Live

Personalized AI storybooks where a child becomes the hero. Full commerce app: characters, story series, leads, Stripe checkout, admin, and narrated video exports in YouTube, Shorts, and square formats.

AI roleLLM story generation, AI illustration generation, and AI voice synthesis rendering finished narrated videos through a Python pipeline.

storyhero.bizinabox.online

Rank & Rent + Voice AI

Live, autopilot

13 local lead-generation microsites (10 live) built by a config-driven Python generator, with call tracking and a written 10-stage pipeline an AI agent can run cold.

AI roleAn AI voice agent is provisioned per site by API, so an AI answers inbound calls for every property. Market validation runs on SERP and keyword APIs with cached research.

brandonradon.ca · lethbridgefoundation.ca · +8 more

HyperFrames Pipeline

Working, parked

Turns raw screen recordings into branded, captioned, auto-zoomed vertical and widescreen videos with no video editor. Evaluated three commercial tools and beat them, so per-video cost is zero.

AI roleLocal Whisper transcription drives word-pop captions; a custom mouse tracker records cursor position and derives zoom keyframes automatically.

Python + FFmpeg, one-click desktop app

Deal Evaluator

In validation

Paste any US auction URL and an agent returns BUY / WATCH / PASS with a maximum bid, for importing liquidation inventory into Canada.

AI roleThe pattern I teach: LLM for research and judgment (parsing lots, finding sold comps), deterministic Python for every dollar of landed-cost math. The model never does arithmetic.

Every evaluation logged; go/kill date set in advance

SpareSquad

Live, parked

A production iOS / Android / web app matching short-handed rec sports teams with spare players: auth, geosearch, swipe deck, chat, ratings, invite codes. Submitted to both app stores.

AI roleBuilt end-to-end with AI-assisted development in weeks; its league outreach campaign was drafted and run by Proof Engine's AI executors against 58 researched organizations.

sparesquad.com · app.sparesquad.com

Free Site Play

Live motion

A lead engine where field VAs find prospects and AI builds every website from an instrumented template with analytics, form relay, and a genuinely enforced 7-day trial expiry.

AI roleClaude is formally one of the three roles in the playbook. Form submissions are AI-summarized before reaching the owner; AI drafts all follow-up outreach.

Sites gated by a live status API, no redeploy needed

Manga Motion

Personal use

Turns comic pages into animated click-through readers using three coordinated model roles: one system, three different AI modalities.

AI roleA vision model reads each page, an image model cleans the art, a video model animates it. Duplicate scenes are detected and reuse cached clips, cutting generation spend.

Cost confirmation before any paid generation

Voice-to-Video

Working

Record a narration once and get a finished slide video: no timeline editing, no manual syncing.

AI roleWhisper transcribes the take, then words are automatically aligned to slides to drive the video timing, with half-second nudge controls for the edge cases.

Python + FFmpeg desktop app

fb-cli

Built, read-only

The governance showpiece: a safe "hands" layer for agent-driven browsing, with every guardrail enforced in code, not in policy.

AI roleRate limits (40/hour, 150/day), human-paced random delays, enforced active hours, a circuit breaker that halts on any checkpoint or captcha, and read-only by default. This is how I think about giving agents access to anything.

Node + Playwright

mini-analytics

Live

Self-hosted analytics with one universal tag on every property I have ever shipped. Its dead-app detector makes "installed but silent" a visible state, so nothing fails quietly.

AI roleNo AI, by choice. This is the measurement substrate every AI project reports into. You cannot manage an AI portfolio you cannot measure.

Privacy-first: no cookies, no fingerprinting, no PII

AI services fleet

Mixed status

Standalone AI services on the same server, built as experiments and utilities: a Whisper transcription API, text-to-speech, a podcast generator, a personalized image generator, a live translator, a website generator, an AI sales chatbot.

AI roleSome run, some are dead containers, and the dashboard says which. An honest lab has failed experiments on the shelf.

All self-hosted on Coolify

Commercial proof

Businesses pay for this work today

These are not portfolio pieces. They are paying relationships where AI and automation carry the fulfilment.

Fractional CTO, US telehealth practice

Active retainers

Two monthly retainers totalling $2,500 USD/month: clinic systems support and a telehealth platform build-out. Integrating a clinical platform with their CRM so patient communications go out reliably, every time. A patient web + mobile app is quoted as a separate $15,000 project.

AI judgment callI ruled OUT using an LLM for patient-facing insight text: it would mean PHI leaving their infrastructure, a BAA with the model provider, and no guarantee against drifting into an outcome claim. Knowing where AI must not go is the senior half of the job.

Invoice automation, barter trade network

In production

Built and operate the member, ledger, and trade-pipeline system for a Winnipeg barter network, now becoming ParseFlow's first AI invoice-parsing tenant. Their bookkeeping arrives as forwarded emails and comes out as structured, validated records.

AI roleVision-model extraction with deterministic validation, human review queue, and per-tenant spot-check rates.

E-commerce + review automation clients

Active

A product company's storefront launched on a CRM commerce stack with the entire 13-product catalogue, shipping, and tax setup scripted and idempotent via API. Separately, a review-automation system captures bad reviews privately and routes good ones to Google.

AI roleAI-built sites and assets; deterministic, re-runnable automation for anything touching client data. Dry-run and verify modes on every script.

What the builds taught me

The judgment layer: what I actually bring to an AI Director seat

Anyone can call a model API. The value is in the operating decisions. Every principle below was earned on a real system, and I can show the code and the written decision for each.

Governance

Approval gates are architecture, not policy

In Proof Engine, AI output physically cannot go outbound without approval: the check lives server-side at the dispatch layer, and the UI check is documented as convenience only. Agent autonomy is a per-agent toggle, off by default. This is the exact pattern an enterprise needs before AI touches customers.

Reliability

Prompts are not controls

My voice system proved that a rule in the system prompt does not reliably hold: the model reproduced a banned phrasing anyway. The fix that held was defense in depth: a deterministic detector plus forcing the model to enumerate violations before rewriting. I design AI systems assuming the prompt will fail.

Risk

Know where AI must not go

For a healthcare client I kept patient-facing text template-driven instead of LLM-generated: PHI residency, BAA exposure, and uncontrollable claim drift outweighed the convenience. The AI Director's first job is drawing that line before someone crosses it.

Cost

Spend tokens like money, because they are

ParseFlow originally ran three model passes per invoice. Measurement showed extraction was already accurate and the extra passes tripled cost while flagging non-issues, so I replaced them with free arithmetic checks plus a 10% random human spot check. Elsewhere: cached clips, deduped scenes, math in code instead of in the model.

Truth

Verify what the model claims

An AI research pass produced a plausible statistic about a prospect that turned out to be false on re-check; it was removed and the catch was documented. Separately, an outreach agent once invented a credential, so I added a truth block and a deterministic detector. Hallucination management is process, not hope.

Portfolio

Every initiative carries kill criteria

Each project gets a written go/kill threshold before it starts, and a decision log records every kill, park, and pivot with reasoning. In one quarter I archived two projects, parked one, and promoted one, against a specific revenue target. AI adoption programs die from zombie pilots; mine cannot.

Full transparency

What failed, and what it taught me

I keep an append-only decision log. These are real entries, not spin. I trust a leader who can show their failures more than one who claims none, so here are mine.

Land flip validationKilled early

A 30-day validation of a land-flipping model. The AI stack was real and substantial: a conversational SMS agent trained on the brand's website, quiet-hours and pacing guards, 280 priced offers queued against 2,887 researched listings. I killed it three weeks before its own deadline.

LessonThe deal shape was wrong, and AI at scale does not fix a wrong strategy. Kill on reasoning, not on the calendar. All work archived, nothing deleted.

Three-pass AI quality controlRemoved

ParseFlow's original design triple-checked every invoice with extra model passes. Real-invoice testing showed the redundancy tripled cost and mostly objected to fields that were not even in the schema.

LessonMeasure before you trust redundancy. Deterministic validation beats stacked model calls. The heavier functions remain in the codebase, tested, in case volume ever justifies them.

SpareSquad partnershipSplit, relaunched

The original app was built under a partnership that split, with the partner keeping the brand and domain. I rebranded, repurchased domains, rebuilt the presence, and resubmitted to both app stores within days. The app is live; growth work is deliberately parked because it does not serve the current revenue focus.

LessonShip speed is a recovery tool. Owning your own infrastructure made the relaunch a rename, not a rebuild.

Self-maintained MCP integration serverDeprecated

I ran a community-built integration server giving AI agents full API access to my CRM. When the vendor shipped an official endpoint, I archived mine the same week with an explicit "do not resurrect" note and a documented decision rule for when to use which integration path.

LessonDo not maintain infrastructure the vendor now ships. Sunk cost is not a roadmap.

Downmarket positioningSuperseded

An earlier $49/month mass-market offer was deliberately superseded by senior fractional-CTO retainers at 30 to 100 times the price point, because the delivery capability had outgrown the positioning.

LessonReposition when the evidence says you are underpriced. The archived document and the reasoning are both preserved.

The foundation

Infrastructure I own and operate

Everything above runs on infrastructure I administer myself. That matters for an AI Director role: I know what these systems cost, how they fail, and how to keep data private, because I operate the full stack.

Layer	What runs there	Why it matters
Coolify (self-hosted PaaS)	78 applications: every project above plus client sites, monitoring, and utilities. Push-to-deploy from private repos.	Full-stack ownership: Docker, reverse proxying, SSL, deploy pipelines, capacity planning.
n8n workflow automation	Webhooks for email, Telegram alerts, scheduled reminders, image generation pipelines.	The glue layer between AI systems and the humans they report to.
Model access layer	Anthropic API direct (Claude Opus vision) and OpenRouter (Claude Sonnet, Gemini image models), swappable per system.	Multi-model fluency with per-task model selection and cost tracking.
MCP integrations	Official CRM MCP server wired into AI coding agents, spanning 43 sub-accounts on one grant.	Hands-on with the emerging standard for giving AI agents safe tool access.
Data layer	Postgres and SQLite per app, self-hosted Git (Gitea) alongside GitHub, uptime monitoring, universal first-party analytics.	Data residency and privacy by default: client photos, PHI-adjacent data, and analytics never leave infrastructure I control.
Agent toolkit	53 packaged AI skills: 11 custom-built automations (content pipelines, deal evaluation, deployment ops, client management) plus a 42-skill marketing pack.	Reusable, documented AI capabilities. This is what AI enablement for a team looks like in practice.

The role

What I would do as your AI Director

The playbook below is not theoretical. It is the operating loop I already run across my own portfolio, applied to your organization.

01 · Research

Map the real workflows

Sit with each department and find where hours actually go. The highest-leverage AI wins are rarely where the hype points; they are in the repetitive, judgment-light work people already hate.

02 · Build

Ship governed pilots fast

Working software in weeks, not decks in quarters. Every pilot ships with a human approval gate, measurement wired in from day one, and a written go/kill threshold.

03 · Govern

Draw the hard lines early

Where data can and cannot go, what AI may and may not say, what runs autonomously and what never will. Enforced in architecture, not in a policy PDF.

04 · Scale

Grow a capability registry

Wins become reusable, documented skills the whole team and its AI agents can use, the same registry pattern my own platform runs on. Kill what does not perform, publicly and on the record.

Prepared by Chibuzor Sam Alumba · August 2026 · every claim in this document is backed by a live system, a repository, or a dated decision-log entry, and I am happy to walk through any of them.