AI you can check.

Bounded systems, not prompt wrappers.

Anyone can call a model. The engineering is in what happens next: what the model is allowed to touch, what stops it inventing an answer, and how anyone outside the building can verify the output. Four AI systems run in production here, and every one of them is constrained by code that can reject what the model returns.

Start an AI project See the guardrails

The Pigskin AI rankings board showing per-player verdicts generated from real formula inputs

AI systems in production4 stacks

Model database access7 tools, no raw SQL

Published outputSHA-256 addressed

Vision inferenceRuns offline

Platform tests1,000+ functions

01The real problem

The demo is easy. The failure mode is the job.

Most AI features work beautifully until the model is confidently wrong in front of a customer. That is not a prompt problem. It is a systems problem, and it is solved before the model is ever called.

The pattern used here is the same every time. Decide what must be deterministic and keep the model out of it. Give the model a narrow, named set of tools instead of open access. Check its answer against a budget after it returns, and reject it if it overreaches. Then publish in a form that lets someone else confirm nothing changed behind their back.

It is worth being direct about scope: every system described here runs in a product or property this studio owns. That is exactly why it can be shown at this level of detail, down to the enforcement code and the published checksums.

For the website-facing side of this work, the SEO, GEO, and AI website field guide shows how public page content, structured facts, and a bounded assistant work as separate layers.

LLM integration with guardrails
Applied machine learning
Computer vision
Governed content generation
Verifiable data feeds
Cost and usage controls

Design ruleThe model never decides alone

AccessNamed tools, not open queries

BudgetChecked after the response

FailureVisible, never silent

Systems review What gets built

Six things that keep an AI system honest.

These are the parts that decide whether an AI feature survives contact with real users. Each one is running in something shipped.

Sys 01

Models on a leash

The useful question is not which model you used. It is what the model is allowed to do when it is wrong.

No open database accessThe analyst gets seven parameterized tools instead of raw SQL. An unknown tool name raises rather than runs
Schema containmentA nine-table curated allowlist against a twenty-five table blocklist; a query outside the contract is rejected
Enforced movement budgetThe model may request bounded exceptions only, and the request is checked after it returns: no more than 15% of a board, and twelve total ranks of movement
Structured outputTemperature 0.15 with a JSON response type, a fixed vocabulary of fourteen adjustment codes, and a recovery path when the response arrives fenced

Sys 02

Output you can audit

An AI answer that cannot be checked is a liability. The rankings feed was built so a stranger can verify it without asking us anything.

Content-addressed publishingEvery published board is named by its own SHA-256 digest, so the filename is the checksum
Create-only uploadsObjects are written with a precondition that fails if the name already exists, so a published board cannot be quietly rewritten
Declared provenanceEach board records the model and prompt version behind every position, including the literal value deterministic-no-llm where no model was involved
Hash-verified consumersThe site that renders the feed compares the digest, downloads only what changed, and verifies the bytes it received

Sys 03

Answers pinned to numbers

The fastest way to stop a language model inventing things is to stop asking it to remember. Every explanation is generated from the record in front of it.

Function calling into the warehouseAnalytical claims are backed by a real query, and a failed query stops the answer instead of guessing
Explanations from real inputsPublished verdicts cite the same touches, target share, and efficiency values that produced the rank
Deterministic orderingPosition formulas decide the order. The model describes the result and, within caps, flags exceptions
Restricted model typesThe warehouse contract allows linear and logistic regression and blocks opaque model types outright, so a ranking cannot silently become unexplainable

Sys 04

Inference on your machine

Not every AI feature should leave the building. VanityFilter is a commercial product whose vision pipeline never touches a network.

Models inside the installerThree pretrained models totaling roughly 20 MB ship in the package and run locally
478 facial landmarksFace landmarking across up to five faces per frame, with confidence thresholds lowered from the defaults because the defaults dropped turned heads
A second detector for the hard casesA separate ONNX face detector catches the profile faces the landmarker returns nothing for
No account, no upload, no telemetryThe product is specified and sold as requiring no internet connection to process a frame

Sys 05

Generation under an editorial contract

Generated copy at scale is a legal surface, not just a writing task. The rulebook is enforced as a system instruction, not suggested in a prompt.

A written forbidden-claims listThe rulebook bans invented prices, promised timings, and claims of owning or operating equipment the client does not own
Visible failureA missing field writes a literal marker into the output rather than leaving stale text that looks finished
Retry with backoffThree attempts at escalating delays on rate limits and transient errors, with a stop control the operator can hit
Length-bounded rewritesGenerated replacements are held inside a percentage window of the original so a layout cannot be broken by a long answer

Sys 06

How the work is run

AI speeds up production only when the contract comes first. This is the method behind every build here, not a slide.

Contract-first designModule signatures agreed in writing before implementation, so parallel work composes without rework
AI-assisted, human-directedThe model accelerates implementation against an agreed contract; a person owns the decision and the review
Measured, not guessedOptimizations are benchmarked before shipping and the rejected experiments are kept with their numbers
Tested as a systemOver a thousand test functions guard the ranking platform, covering ingestion, formulas, guardrails, and publication

02Where it runs

Four stacks, four different problems.

These are deliberately unlike each other. One is a cloud data platform, one is desktop computer vision with no network at all, one is content generation at scale, and one is production imagery.

Pigskin AI ranks players with formulas trained in a data warehouse, then has a language model explain each rank from the same inputs that produced it, and publishes the result as a checksum-addressed feed anyone can verify. VanityFilter runs face landmarking and segmentation entirely on the customer machine, with the models bundled in the installer. SKYNET generated localized copy for more than 60 live websites under a written rulebook that forbids the claims which create legal exposure. American Heavy Lift Helicopters took its scene imagery from a first-party generation pipeline with a second provider held in reserve.

Applied MLWarehouse formulas, backtested

Computer visionOffline, on the customer machine

Generation60+ sites under one rulebook

ImageryTwo providers, no single point of failure

Straight answers.

The questions worth asking anyone who says they do AI.

How do I know the AI is not making things up?

Because it is not allowed to answer from memory. In the ranking platform the analyst has no open database access at all: it receives seven parameterized tools over a curated allowlist, raw tables are explicitly blocked, and a failed query stops the answer rather than letting the model improvise around it.

Does the AI make the decision, or describe one?

On the ranking product it describes. The order comes from deterministic position formulas, and the published provenance field literally reads deterministic-no-llm. The model may request bounded exceptions, and its budget is enforced in code after the response returns.

Can you build AI that does not send our data to a third party?

Yes, and it is already sold as a product. VanityFilter bundles three pretrained models inside the installer and runs its vision pipeline entirely on the customer machine, with no account and no upload. Whether that fits your case depends on the work, but it is a shipped pattern here rather than a theory.

What happens when the model fails?

It fails visibly. The generation pipeline retries with escalating backoff, then writes an explicit failure marker into the field instead of leaving text that looks finished. The publisher fails closed: if a board is incomplete it does not publish.

Have you delivered AI for an outside client yet?

Not yet. Every AI system described here runs in a product or property this studio owns and operates, which is why it can be inspected in this much detail. The AI-assisted build work has run on a client program: a network of more than 60 localized sites generated under an editorial rulebook.

Which providers and stacks have you shipped on?

Google Cloud end to end for the ranking platform, including the warehouse, scheduled jobs, and storage. Gemini is the language model in production across the analyst, the site factory, and rank explanations. The imagery tooling calls the OpenAI Images API as its primary path with a second provider as backup, so a single outage cannot stall a build.

Have a problem an AI system might solve?

Bring the workflow that is costing you time. We will tell you honestly whether it needs a model at all, and what it takes to make the answer trustworthy if it does.

Open project channel