I'm Clayton Camera — a technical founder who builds AI systems, autonomous agents, and full-stack software that find hidden revenue and automate the work teams shouldn't be doing. Below is a selection of what I've shipped — live products, not slides.
*Author-curated coverage index for the BTC research workspace — a breadth measure, not a benchmark or accuracy score.
🛠 Work With Clayton
I build the same kind of systems you see below — for companies. Every engagement starts with a fixed-fee audit that puts a dollar figure next to every leak, ranked by speed-to-cash. The audit credits toward the build if you move forward, so the diagnosis costs nothing if we ship together.
Portfolio
Real ventures and systems — live products and deep research. Each one lists who it helps and the edge it delivers.
The routing layer for the orbital-compute era.
Decides where an AI workload should run — orbital GPU capacity or terrestrial cloud — using real orbital mechanics: SGP4 tracking across 12,000+ satellites drawn from the official 32,308-object catalog, 37 modeled ground stations, eclipse and weather physics, 23 autonomous engines, 5 USPTO filings. Its cost-compare API (free key) still returns a "use terrestrial" recommendation today and names the exact rate at which that flips — the platform tells you when not to buy it. Rebuilt on that honest footing in July 2026 after a 30-agent adversarial audit of its own claims. The AI Compute Audit went purchasable in August — three tiers, live checkout, fulfilment deliberately by hand.
The operational backbone of a live peptide brand.
A live WooCommerce brand where I'm CTO & Managing Partner. I architected Command Center 2 (CC2) — the ops plugin running shipping, payment verification, inventory, COGS, email, and order workflows — plus the affiliate system behind 37 active affiliates, a customer rewards engine (credits earned per order, redeemable at checkout), and a promotion engine that arms tiered site-wide sales and bundle stacks on schedule without anyone touching the store at midnight. Its first full run — the July Summer Sale — closed at $6,254.70 across 34 orders and disarmed itself on time; its second, the Back to School Sale, armed itself on schedule Aug 22 and runs through Sep 7 — 46 products tiered 10/20/30% by all-time units sold. In August the brand itself moved: 95 product images were rebuilt so the whole catalogue carries the current label, by replacing the label alone rather than upscaling a photograph that would have degraded every render it touched.
A paper-first control plane for multi-algorithm trading research.
A Money Making Solutions research stack: real-data ingestion, strategy experiments, bot orchestration, risk gates, paper ledgers, and explicit promote-or-kill decisions. Brokerage execution is locked by default behind hard caps. In August 2026 all nine strategies were put through a 22-agent audit where every claim had to survive a second, adversarial reviewer. The verdict was that none of them are investor-ready — the best venue win rate on the authority ledger is 56.5%, and the honest sustained throughput is about two trades a day. The public preview shows the architecture and evidence without exposing the owner controls.
Momentum intelligence across five worlds at once.
A multi-domain AI engine that scores, ranks, and tracks entities across sports, markets, brands, politics, and entertainment. Narrative ingestion, Polygon.io market data, AI sentiment scoring, and alerting over 90 entities across 5 domains. The API has been up continuously for a month; the ingestion worker is currently idle rather than pulling new material.
A 15-chapter Bitcoin research workspace.
Covers macro cycles, on-chain metrics, technicals, sentiment, institutional flows, trading frameworks, and risk management — with live market data, a 3D diagram of the system, and a research activity feed. It also keeps a public forecast ledger: every prediction is timestamped, resolved on schedule, and scored against a dumb baseline. 3,503 resolved forecasts in, the global record is still 6.4 points behind always-guessing-up — and the page says so. The pooled 24h slice is positive after the maker-fee assumption, and on Aug 23 the guard24 candidate cleared the promotion gate for the first time — 190 resolved, 65.8% hit, Wilson lower bound 58.8% — though it stays paper-only pending manual review, with the global record still behind. Coverage Score: 189* (author-curated breadth index, not an accuracy score). Educational research only — not financial advice.
A vault where the server never sees your data.
A zero-knowledge encrypted vault for your most sensitive documents. AES-256-GCM encryption runs entirely in your browser — the server never sees your data. Store IDs, insurance, financial accounts, medical records, passwords, and estate documents with military-grade security.
Autonomy that can't act without your approval.
A local-first autonomous AI agent with a non-negotiable human-in-the-loop gate on every side effect. It reads, plans, and proposes — but every trade, file write, deploy, and API call waits for an explicit human APPROVE. Zero OpenAI dependency: Ollama locally, Anthropic or Perplexity in the cloud. 30/30 adversarial tests green, hash-chained audit log, kill switch.
Scattered peptide research, made legible.
A research-grade peptide knowledge hub — 79 structured entries across 10 research categories, plus 119 programmatic side-by-side comparison pages, a reconstitution calculator, a 77-term glossary, and a research library: 264 crawlable URLs in total. Every compound plate is generated from the entry's own data, so a picture can't contradict the record it illustrates. Educational research only — not medical advice, and PeptiDex does not sell peptides.
A 30-second door into door-to-door sales work.
A recruiting platform for door-to-door fiber sales, now carrying three carriers — T-Mobile, Frontier, and Kinetic. 294 city pages across 47 state and metro hubs, a bilingual (English/Spanish) apply lane, and a 30-second application with no résumé. The August expansion came from one join: ranking markets by coverage × population instead of coverage alone, which surfaced the largest uncovered Frontier market in the country sitting at roughly position 1,100 on the old list. Machine-readable job feeds are earnings-claim-free by design and guard-tested on every deploy.
Catches mispriced lines before the market does.
A 12-module NBA betting intelligence system built to detect line mistakes before the market corrects. Real-time odds modeling, schedule-stress analysis, pace projections, injury-impact scoring, prop-market edges, an arbitrage window scanner, and CLV tracking — all in one dashboard.
A performance-nutrition brand, built and ready.
A performance-nutrition brand concept — fully built and researched end-to-end, not yet for sale. Powered by one of the most protein-dense organisms on earth: a target of 28g complete protein per serving, all 9 essential amino acids, and a ~95% lower carbon footprint than beef. Brand, product spec, and storefront are built and ready for launch.
The West's foundational thinkers, source-traceable.
A public-domain reading room for the philosophers who built the operating system of the West. 16 seed authors — Plato, Aristotle, Socrates, Heraclitus, Parmenides and more — with a compact catalog, source-traceable provenance (Gutenberg, Open Library, Wikidata, Perseus), and an in-browser reader. Modern copyrighted translations are never bundled.
Nothing ships until it clears six gates.
A full-stack viral-content command center for the LoopholeMaxing studio. Plan, draft, and ship content through a six-gate pre-publish workflow, with a content library, hook-variant generator, Trend Radar, Idea Inbox, deterministic scoring, and an audit log. Built on React + Express + Drizzle + Postgres. Owner-gated — sign-in required.
Import, track, trim and export — in the browser.
A personal content command center. Import footage from GoPro, Nikon, and phone — Studio auto-tags where every clip came from, tracks it through a Raw → Editing → Clipped → Posted pipeline, and lets you trim and export real clips right in the browser (ffmpeg.wasm). A companion auto-clip engine runs on my Mac mini: drop raw footage in a watch folder and it returns hardware-encoded vertical cuts with on-device transcribed captions, synced back to every device. Multi-device sync and invite-based multi-user shipped July 2026.
Every public camera in one live map.
A real-source public-camera aggregation dashboard for Tallahassee and Leon County. 407 verified cameras dedup'd from 593 raw feeds across City ATMS, FDOT, FL511, WeatherSTEM, FAA, and regional webcams — live snapshots proxied through a per-source Referer-aware backend. Map-first, dark-themed, with live traffic incidents and congestion overlays. Its deploy has been down since 7 June 2026 and hasn't been prioritised back up; the code and the camera ledger are intact. Listed here as built work, not as something you can open today.
Ship Log
The systems above aren't finished products sitting still — they get worked on every week. This is the last 30 days, dated. Every line landed on a live system, and the ones where the answer came back negative are here too.
The resource contract is now modality-aware — CPU/GPU/NPU/FPGA/QPU — instead of GPU-shaped, with two rules enforced in code rather than policy: a request that states nothing stays GPU-only, so new hardware can never silently capture existing traffic, and capability is evidence-gated — a simulator profile is listed but never compute-eligible. Four public read-only endpoints shipped with it, including a deterministic benchmark anyone can run: the reference solver honestly scores 0 wins and 20 ties against the classical baseline, and the page says so.
Honesty wave on the fleet itself. The public discovered-satellite count was quietly inflated — 27 of 264 rows were objects the catalogue declines to name, and the endpoint now reads the honest figure. The same wave found the fleet cap breached in a way that could never self-heal: a cycle starting at 11,998 ran its full batch and parked prod permanently red at 12,048 — fixed with cap enforcement inside the add loop and reversible soft deletes. Placeholder "TO BE ASSIGNED" catalog names no longer publish as satellite identities. Canary back to green, now 15 checks.
An accuracy pass over every affiliate-facing surface, because you cannot redesign a dashboard while it gives three different answers to "what do I earn?". The public tier ladder advertised a 50% personal discount the engine has never granted — the same page said 30% higher up, and the wrong number was the more generous one. Live copy now matches the engine everywhere, and the six Elite affiliates finally see their real 15%/16% rates instead of a generic figure.
The satellite catalog query had been stopping at NORAD id 100,000 — and the five-digit catalog was exhausted in July, so every newer object was structurally invisible. Extended to full legal Alpha-5 scope: the census is complete at 32,306 catalog objects, highest identifier 270,449, with a fifteenth canary check that proves full-scope publication on every cycle instead of trusting it once.
The promotion gate cleared for the first time: the guard24 24-hour candidate reached 190 resolved forecasts at a 65.8% win rate with a Wilson 95% lower bound of 58.8% — above the bar — and positive expectancy after maker fees, so the signal contract flipped from shadow to actionable_paper. Same scoreboard, same rules that said no for months. Still paper-only, still zero trades: manual promotion review is the next gate, and the global ledger remains behind always-guessing-up.
The promotion engine's second full run armed itself the night before and opened on schedule at midnight: the Back to School Sale, up to 30% off through Sep 7, tiers assigned by all-time units sold and generated from live order data by script rather than hand-picked. It registers its own fee-label prefix and purges cache at both ends — the two traps the Summer Sale taught. Same wave, CC2 closed the self-referral loophole: an affiliate buying through their own link no longer earns commission on their own purchase.
A $374 affiliate order went out the door for $47.51 — her own coupon, the 30% personal discount, and peptide credits all stacked, for a true P&L of minus $100.01. Shipped a 50% total-discount ceiling across the whole money path: every discount source still applies, credits calculate last and absorb the clamp, so no combination of promotions can ever discount an order past half price.
Ten dictated storefront edits shipped live: two stacks renamed with 301s, savings rendered red sitewide, checkout instructions moved onto a high-contrast panel, the footer gmail replaced with the brand address, and an unverifiable US-manufacturing claim deleted. Two of the edits briefly took the site down — both my own mistakes, both reverted from backup within a minute, both on the record.
Rebuilt the research stacks from the order book. Nine stacks had sold exactly one unit ever — 0.4% of all-time revenue — and market-basket analysis showed why: five of nine recipes contained a SKU nobody buys, and none included bacteriostatic water, which attaches to 15.6% of orders on its own. The new stacks are built from combinations customers actually assemble. Same wave: the mobile category page had been spending ~700 of an 812-pixel screen on chrome before the first product — the grid now shows 163px sooner.
Continue with Google is live on the storefront. The case was measurable, not cosmetic: 60 guest buyers have no account, and 260 earned credits sit unredeemable because redemption is login-only — sign-in is the lowest-friction path that converts those buyers into account holders.
Expansion waves 17–21 in four days: West Texas bilingual, Washington and Wisconsin beachheads, in-turf gap markets, the I-75 corridor, and two Spanish-primary funnels — the surface grew from 259 cities to 294 cities across 47 hubs. A routing fix in the same wave stopped expansion markets from swallowing active-turf area-code pins.
Turned the trading estate into two honest, connected surfaces: a public MMS research preview and a full owner Trading Cockpit inside The Window. The public page shows the architecture, all promotion gates, the adverse audit result, and a live pull from the BTC ledger; the owner view maps the complete bot roster, source-of-truth hierarchy, open risk gaps, and rebuild order. Nothing was promoted or enabled. The live candidate remains shadow-only, and every unavailable feed fails closed instead of manufacturing a green dashboard.
Audited the outreach database and found the thing being sold had been pitched in about ninety-nine emails and linked in none of them — every send pointed at a status page instead of the buy page. Checkout attempts to date: zero, which is exactly what you would expect. The cold-email experiment running on top of this had no buy path in it, so a null result would have been read as "cold outreach doesn't work" when the ask had never actually been made. Fixed the template. The drafts already approved and queued still carry the old body, because the scheduler sends stored HTML verbatim — and rewriting an approved message would quietly invalidate the approval that made it sendable, so that one is a decision, not a deploy.
The application form's spam honeypot was a hidden field named company. That is also a Chrome autofill token — so the browser was filling it in for real applicants and the form was silently discarding their applications as bot traffic. Nobody complains about this one, because the people it affects never know they didn't apply. Dropped submissions are now logged instead of vanishing. Same wave: the post-submit view was scrolling away from the "save this number" step on every converting lead, and the Google Jobs markup was violating Google's own HTML-format rules site-wide. 2,018 tests green.
An audit in July disproved a performance number that was still sitting on my own sales page. The fix was written and committed on Aug 6 — and the static-site build failed in the same minute, silently, so the live page kept serving the disproved claim for two more days while git showed it resolved. Caught by diffing the deployed HTML against the repository rather than trusting the commit. All pages now match byte-for-byte. "Shipped" means the bytes changed at the URL — a merged commit is not evidence of either.
Retired the old cartoon mascot from the entire catalogue — 95 images, and zero products, variations or categories still carry it. The new brand existed only as marketing photography, whose largest usable vial crop was 285×766 against a live catalogue of 2000×2000 renders with a 608×717 label. Rebranding by upscaling that photo would have been a sevenfold enlargement of every element including the type: downgrading fifty-six products to make them match a brand. So the vial was kept and only the label replaced — the glass, cap, lighting and reflections in the existing renders carry no branding, only the label does. The two elements that are the mark were cut as real pixels from the source. Nothing was redrawn by a model.
A compound rename swept every line of text on the store — and left the old spelling baked into the pixels of a product photo for sixteen days. Rebuilt the composite from its original sources, proved the rebuild was byte-identical to the live file before touching anything, then swapped it under a new filename so no CDN could serve the old one.
Put all nine trading strategies through a 22-agent audit in which every finding had to survive a second agent whose job was to refute it. The verdict was that none of them are investor-ready — the best genuine win rate on the entire board is ~60.5%, and honest sustained throughput is about two trades a day. Published the negative result rather than the backtest that would have read better.
Every market ranking until this point sorted on carrier coverage percentage, which put towns of four hundred people at the top. Re-ranked on coverage × population — one join against census estimates — and the largest uncovered market in the country turned out to be sitting near position 1,100. Shipped 30 city pages and 5 metro hubs off that finding; the surface is now 255 cities across 44 hubs.
The orbital-elements feed cut us off, and the physics layer went blind — pass windows collapsed from 102,000 to zero. Moved the platform onto the official 31,672-object catalog as its primary source. The rule that came out of the outage: a 304 Not Modified means our copy is current and can be trusted; a timeout means nothing, and relaying it would stamp unknown-age data as fresh.
The AI Compute Audit became purchasable for the first time — three live price tiers and a real checkout session. Fulfilment is deliberately manual: a sale pages a human rather than triggering a generator, because the deliverable is the part that has to be right.
Ran 64 queued outreach drafts past a 7-agent pre-send review before a single one left the building. Seven were held: a stale fleet count, two notes greeting the wrong person at the right company, and one addressed to a firm that had entered receivership. 57 went out.
Moved the money-path tests into CI. Commission math, credit ledger, and discount arbitration are now checked on every push instead of by hand after a release — which is where they used to get checked, when someone remembered.
The Summer Sale ended on schedule and disarmed itself — tiered discounts and nine bundle stacks off at the stroke, no midnight store edits at either end. Final: $6,254.70 across 34 orders.
Five independent agents went through the live site hunting for claims that couldn't survive a curl. What they found got corrected, not softened — the same standard the July audit set.
All 79 compound images had come from one function that drew four shapes and recoloured them by category — so for roughly thirty compounds the picture contradicted the entry's own data: non-peptides drawn as injectable vials, topicals given an injection route, two different compounds rendered byte-identically. Replaced with plates derived from the real fields, gated so a residue strip is drawn only when the sequence actually matches the stated length.
Wired five domains onto one person entity — a single id, defined once, referenced everywhere — then pointed 30 agents at the result and told them to prove each other wrong. Two already-published claims came back false and got visible corrections rather than quiet edits.
Updated as things ship. Internal ops tooling is deliberately left off this page.
That's the whole job. Start with a 48-hour audit that puts a dollar figure next to every leak in your business — the diagnosis credits toward the build if we ship together.
Work With Clayton →📬 The Loophole Letter
Build logs, revenue experiments, and the systems behind everything on this page — straight from the operator, not a content team. Free, no spam, unsubscribe anytime.
Your email goes straight to my own CRM. No third-party lists.