Work

A production AI platform, and the shape of the work.

The system we built and operate ourselves — multi-tenant, six model providers, live billing — and, by sector, what a typical engagement involves: the volumes at which it makes sense, what gets built, and what changes once it is running.

Built and operated by us

Our own product, in production.

Our own product. Live, multi-tenant, with paying customers — and the reason we can talk about production AI from experience rather than from a whitepaper.

Real · oursApplied AI SaaS

We built and run a multi-tenant AI platform that measures brand visibility across six AI engines.

The problem

Buyers were shifting their research into ChatGPT, Gemini, Perplexity and AI Overviews, and no brand could see whether it was being recommended, ignored or misdescribed inside those answers. Measuring it needed a platform rather than an opinion: six providers, per-tenant isolation, and a cost model that survives running thousands of queries a day.

Multi-tenant AIModel orchestrationCost governanceAI visibility

What we built

Multi-tenant SaaS on Next.js and Postgres with row-level tenant isolation
Orchestration across six model providers with per-tenant cost ceilings
Citation and share-of-voice extraction from live engine responses
Accuracy monitoring that detects when engines state something false about a brand
Automated reporting, billing and multi-region delivery

Where it stands

Live with paying customers across brands, agencies and public-sector mandates — and the reason our AI visibility service costs us less to deliver than anyone reselling a third-party licence.

We will demo it live on a call. That is a better test than any case study, because you can ask it questions we did not prepare for.

What operating it taught us

Everything we know about production AI, we paid for.

Six things that went wrong on our own platform before they could go wrong on a client’s. Each one is now a standard we build to — which is the difference between a practice and a preference.

A model update broke us before it broke a client

A provider shipped a new version and our extraction quietly started missing citations in a category we had not tested. We found it because a number moved, not because anything errored.

What we do now

Eval suite and regression gate on every build

Context grew and the cost model stopped holding

Per-call cost was never the problem. History accumulated, retrieval returned more chunks, a prompt got extended — and the same work cost several times what it did at launch, on flat traffic.

What we do now

Cost architected up front, ceiling that alerts

A confident wrong answer is worse than no answer

The system produced something plausible rather than nothing, because that is what the objective rewards. For a client that is an invented delivery date, or advice nobody is licensed to give.

What we do now

Citation enforced, refusal designed before the happy path

Tenant isolation has to be structural, not careful

Multi-tenant AI fails in a way ordinary software does not: one tenant’s data reaching another through a cache, a shared index or a prompt. Being careful is not a control.

What we do now

Row-level isolation, per-tenant retrieval boundaries

Nobody reads a dashboard nobody is accountable for

We instrumented accuracy, drift, cost and latency early, then did not look at it for weeks. A daily all-clear gets ignored, and then the one that matters gets ignored too.

What we do now

Named owner, monthly number, alerts only on failure

The demo is the easy part

All of the above happened after launch. The build was weeks; the eighteen months since is where the system either stayed accurate, affordable and trusted, or quietly stopped being any of those.

What we do now

We run what we ship, with a monthly report

Engagement patterns

What a typical engagement looks like.

Composites by sector — the volumes at which each build pays, what gets built, and what changes once it runs. Useful for sizing your own before we speak.

Pattern

Multi-location clinic

The situation

Around 900 enquiries a month across WhatsApp and phone, two people answering both.

What we would build

WhatsApp assistant for questions and booking, voice agent for overflow calls after 7pm.

What changes

Enquiries answered within seconds instead of hours, front desk freed for people who are physically present, and every after-hours call captured instead of lost.

Pattern

Real-estate developer

The situation

A campaign generating several thousand leads a month, three salespeople, no way to rank them.

What we would build

Lead qualification and routing on budget, location and timeline, with automated follow-up for the rest.

What changes

Sales time concentrated on the enquiries most likely to convert, and nothing dying of silence because nobody got to it.

Pattern

Manufacturer, accounts payable

The situation

Roughly 4,000 supplier invoices a month, handled by hand, reconciled against purchase orders.

What we would build

Document extraction with validation rules, exceptions routed to a person, clean data into the ERP.

What changes

Typing removed from the process, disputes surfaced earlier, and month-end no longer dependent on one person clearing a backlog.

No invented percentages, no fabricated ROI figures. When we have a client result we are allowed to publish, it will appear above this section with their name on it.

What we build

The capabilities, and we will show you each one working.

Conversational AI

Assistants that talk to your customers and know when to stop.

WhatsApp assistantsVoice agentsWebsite chatMultilingual (EN/HI)

Retrieval & knowledge

Answers grounded in your own documents, with the source attached.

RAG pipelinesHybrid searchCitation enforcementRefusal design

Automation

Multi-step work that runs without a person watching it.

Document extractionWorkflow orchestrationCRM & ERP integrationHuman approval gates

Measurement

Knowing whether any of it actually worked.

Eval harnessesAccuracy monitoringCost telemetryAI visibility tracking

Platforms & integration

The surfaces and systems all of the above plugs into.

Next.js buildsShopify & WordPressCRM / ERP integrationData pipelines

Demand

Making sure somebody actually arrives at what we built.

GEO / AI searchSEOPerformance marketingContent pipelines

In the agreement

Things you can hold us to.

Commitments you can enforce, written into the contract before anything starts.

A fixed price before we start

An accuracy number, not a promise

Model costs at cost, always

A running-cost ceiling

We will tell you not to buy something

A named person who answers

We would rather be a new name that tells you the truth than an old one that borrows somebody else’s. Start with the free scan and judge us on something we actually did for you.

Be the first name on this page.

Our first client AI engagements are priced for the case study rather than the margin. If you have a process worth fixing and you are willing to let us publish the result with your name on it, say so — that is worth a great deal to us and we will price it that way.

Start a conversation Get a free prototype