Case Studies

AI that earns its keep.

Very different domains - our own data-driven creative engine, and a platform we built and run for a live client - held together by one way of working.

Book a discovery callSee our services

A glowing network of connected nodes and data streams - production AI systems at work.

FPI Social Video

One method, pointed at what actually converts

Our own product - and the system that taught us to run agents in production. Not more creative. Creative that performs: the angle is chosen from real performance data, generated only once it is likely to win, and judged on what it returns.

A finished social spot from the engine - hook, product, and call-to-action, generated and assembled end-to-end.

The challenge

Brands do not have a volume problem. They have a hit-rate problem. Which hook lands is unknowable in advance - you find it by testing. But at traditional production cost, testing twenty angles to find one winner is irrational, so most brands guess, ship three, and hope.

What we built

A taxonomy-driven, multi-agent system that runs the full creative pipeline: brand analysis → persona development → storyline → performance-ranked concept → image and video generation. Six-plus autonomous agents (Orchestrator, QA Tester, PM, Fixer, Prompt Optimizer, and more) coordinate the work, with a performance-data layer scoring concepts before they are produced.

How it works

An autonomous QA loop tests its own output, triages issues, and applies safe, bounded fixes - escalating to a human only when needed. Every project feeds compounding learnings at the system, vertical, brand, and product level, so quality improves over time.

The outcome

Enough on-brand angles to actually find the winner, instead of betting the quarter on three. Creative judged on what it returns, not what it cost to make - and a system that gets sharper with every project instead of plateauing. What that has been worth is below, in our clients’ words.

A sample of the output

On-brand creative, produced at volume

Real product and social creative the engine makes across beauty, wellness, supplements, and food & beverage - a sample of the range and caliber, not one hero asset.

Energy-drink creative - electric splash on near-black

Face-cream creative - a golden particle orbit on deep navy

Supplement creative - an explosive protein-powder burst

Coffee creative - a can suspended in a burst of beans

Fragrance creative - a perfume bottle in a violet particle burst

Cosmetics creative - a pressed-powder compact in a rain of gold

The data loop

What makes it data-driven

The generator does not guess. Live results feed a model that scores the next concept, so production goes to what is likely to win - and every result sharpens the score.

Measure

Every asset’s real results - spend, watch-through, engagement, conversion - land in one insights layer instead of scattered across ad accounts and dashboards.

Predict

A ranking model scores each concept by predicted performance, so production goes to the angles most likely to return - and the weak ones never get made.

Generate

Image and video are produced only for concepts that clear the bar, then judged on what they return - and every result retrains the ranker.

Built for five DTC verticals

The same system, retuned per category - so the brand knowledge and the winning-hook data compound where you actually sell.

Sample Beauty product creative from the FPI Social Video engine — Beauty

Sample Supplements product creative from the FPI Social Video engine — Supplements

Sample Food & Beverage product creative from the FPI Social Video engine — Food & Beverage

Sample Fashion product creative from the FPI Social Video engine — Fashion

Sample Wellness product creative from the FPI Social Video engine — Wellness

Days

From brief to first creative, not months

Data-ranked

Every concept scored before production

Self-fixing

QA loop with human escalation

Brands this engine makes creative for

“FPI does all our creatives and they’re the ones that are converting. They took the time to get to know our product and knew exactly what hooks to put in the videos.”

DDr. Greg

“Within a week of upgrading our marketing assets to video, we saw our cost per acquisition drop to half of what it was.”

MMax L.

“They transformed my products into incredibly realistic 3D animations. The team was patient in bringing my vision to life - I’ve since used them for six products.”

CCarole R.

Interconnected records and audit trails as glowing nodes across a dark field

Multiple domains, one playbook: autonomous systems safe enough to trust and good enough to keep.

FPI Operations

The work they delivered but never billed

A real client engagement: FPI built and operates this platform for PTS, a New York City construction special-inspections firm - not an internal demo.

The challenge

Inspection projects across New York City, with crews in the field. The signed proposal sits in a Word file; the work lives in a spreadsheet, updated visit by visit over years. To know whether the two match, someone would have to open every proposal, find the fee table, and count. Nobody ever had - not through carelessness, through arithmetic.

What we built

A 21-skill agentic platform that reads every signed proposal - Google Doc, PDF, or Word - and finds the fee matrix wherever it is buried. Not the line total: the per-visit unit cost, because that is what an extra visit is actually worth. Then it counts what was really performed, and subtracts.

The hard part

Half the visit records name the inspector by initials, and half of those are blank. The obvious fixes failed: file ownership and photo metadata both point at the back-office coordinator who uploaded the batch, not the person on site. We tested them, proved them wrong, and threw them out. What worked was that field crews name their photo folders by date and initials - written by whoever was there.

The outcome

Six figures of work delivered and never billed, surfaced across the portfolio - plus the reporting to keep it from reopening. Unattributed visits down by roughly a third. The firm’s drive was never written to, not once.

Stacks of contracts and ledgers receding into darkness, a thread of light finding one of them.

Six figures

Delivered work found unbilled

~92%

Forensic attribution accuracy

Read-only

Never writes to your system of record

Where the money and the time actually go

One number is easy to quote. The value is that the same reconciliation surfaces eight kinds of opportunity every month - and nobody has to chase any of it.

Work done, never billed

Delivered beyond the signed proposal and never invoiced. Six figures surfaced in the first month - and the sweep is not finished.

Money already owed

Invoices aging across the whole portfolio, surfaced before they turn into write-offs.

Sold, not yet delivered

Visits still remaining under signed proposals. Revenue already agreed, waiting to be scheduled.

Work being given away

Off-scope inspections nobody quoted for. Every one is an upsell conversation that was invisible before.

Projects gone quiet

Anything idle for 45+ days, batched into client outreach - so volume comes back instead of drifting.

Who is busy, who is not

Utilisation by inspector, so more work goes where there is capacity rather than to whoever is already overloaded.

What to send a crew to next

Upcoming visits batched by site and drafted into client outreach - so a crew makes one trip instead of three, and more inspections fit into the same week.

Work with no paperwork behind it

Every visit crosswalked against the report that should exist for it. Missing or mismatched reports surface here, not when a client asks for them.

Each of those used to need a person to go and look, across hundreds of projects, and so nobody ever did. Now it lands in a report every month. The project managers stop chasing hand-offs and start working the list - and the same visibility that finds the money also decides where the crews go next week.

Strip the nouns and none of this is about inspections. A company’s own truth was scattered across systems nobody had ever compared, so nobody could see it. That is every company. This one’s money was sitting between a signed proposal and a spreadsheet. Yours is somewhere else - in fulfilment, ad spend, returns, inventory, or a column someone added in 2021 and never explained.

The method transfers: read what is actually there, compare what has never been compared, say what is true, touch nothing. The specifics never transfer - this firm’s answer was hiding in photo-folder names, and no system built from the outside would have guessed that. Finding your version is the work.

Want a system like these in your business?

Thirty minutes, no pitch deck. Tell us where the manual work hurts most, and we’ll tell you straight whether AI can fix it.

Book a discovery call or email services@fpi.cc