Ten cases, no client names

Problem

Every ad creative, landing page screenshot, listing and user upload has to pass the rules before it goes live: prohibited categories, misleading claims, text hidden in the image, brand and logo misuse, the wrong audience. Doing it by hand does not scale, and sending it all to a cloud API means every creative, every advertiser and every screenshot leaves your perimeter.

Solution

Self-hosted models classify everything that comes in and flag violations. OCR and CLIP are the cheap first stage: they pull the text and visual features off each image and hand them to the classifier as signals, so the heavy language model only sees what needs judgement, such as a claim that reads fine but is illegal in that market. Only the disputed cases reach a person. The rule set is yours and changes without a vendor ticket. The quality bar is measured on a golden set of your own creatives before anything goes live, not taken on the model's word.

Problem

Speech in one language has to be readable as text in another, right away: calls, meetings, support conversations that cannot be sent outside.

Solution

Self-hosted speech recognition and translation, streamed as it happens, with a latency target of a couple of seconds. Models are tuned to the client's domain and vocabulary and run on the client's hardware.

Problem

Years of documents, matters, templates and correspondence that nobody can find at the moment they are needed.

Solution

A private search that understands a question in plain language and answers with links to the original document. Hybrid retrieval: vectors plus full text, then re-ranking. Everything stays inside the company, nothing is indexed outside.

Problem

Running LLMs and vision models in production cheaply, reliably and on your own hardware, without paying per token to an outside API.

Solution

A fleet of GPU servers with request routing across pools, continuous batching, per-model metrics and cost accounting, on our own inference images. The controller that grows and shrinks the fleet is its own case below.

Problem

Invoices, bank statements, contracts and forms are typed into accounting and case systems by hand. It costs hours and produces errors.

Solution

Two tools with two jobs: OCR produces a fast, cheap text layer for the pipeline and the search index, and the vision-language model does the understanding: it pulls the fields out into a structure, parties, amounts, dates, terms, in any of the languages involved. A person confirms instead of retyping. The same engine checks a file for completeness: what is missing, what has expired.

Problem

People forget to log their hours. Work that was actually done never reaches the client's invoice, and the firm quietly loses money every month.

Solution

The system runs on the firm's own servers and looks only at work mailboxes and matter files, under the access policy the firm already has and with the team informed. It decides whether a service was delivered and raises an alert: work done, not logged, not billed. It keeps no message content, only the flag. It fires before the billing cycle closes, so the hours can still be recovered.

Problem

A shared inbox that someone has to sort by hand every morning.

Solution

The system classifies what came in, extracts the substance, routes it to the right person and drafts the routine replies. Nothing is sent without a person approving it.

Problem

Invoices, statements and receipts are entered into the ledger by hand, under strict rules. That is exactly the kind of work automation is good at.

Solution

Reading the document is the same engine as in data extraction. The work here starts after that: a posting-rules engine maps each document to the firm's chart of accounts, VAT treatment and cost centres, checks the entry against the rules, and writes it into the accounting system through its API or import format. Nothing is posted without a person approving it. The rules are strict, so quality is easy to verify.

Problem

Speech synthesis and voice cloning without handing the data to an external service.

Solution

Text-to-speech models on the client's GPU, voice cloning, generation of segments for content or narration. Audio, voices and scripts stay on the client's hardware.

Problem

GPUs are expensive. An idle card eats the whole economic case for owning hardware.

Solution

A controller brings models up and down with the load, mixes in rented GPUs at the peaks and shuts everything down when idle. Scale-to-zero is for batch and peak pools, where a minute of start-up is acceptable; latency-sensitive routes keep a warm minimum, and model weights stay cached on local disk, so a cold start is a planned minute on a rented card, not a timeout for the user. The state of the fleet is readable in one command.