AI and machine learning, deployed

We do not sell «artificial intelligence» as a product. We take a process where someone re-reads or cross-checks the same thing daily and move the repetitive part to a model, leaving the judgement and the veto with the person.

Where a SQL query or a well-configured CRM pipeline would do it, we say so: a neural network there buys cost and one more point of failure.

• LLM assistants: answers from your own procedures, each citing its source
• Computer vision: number plates, people counting, PPE compliance, defect detection
• Forecasting: demand, stock, churn and shift load from your trading history
• Data extraction: invoices, delivery notes and contracts, from scans into fields

A model knows nothing about your internal rules. RAG is a search over your own corpus whose results go into the prompt: documents are chunked, embedded as vectors, stored in pgvector or Qdrant, retrieved by hybrid search — vectors plus classic BM25, because part numbers and surnames are what vector search handles worst.

A cross-encoder then reranks, and five to fifteen chunks reach the model rather than the whole archive. A 128k context window is no invitation to dump everything in: as context grows, so do the bill and the odds the key paragraph is lost in the middle. Locally, scans that mix Uzbek Latin, Cyrillic and Russian need proper OCR first.

Before a line of code, we build a set of 100–300 real questions with reference answers from your support history. Your staff write it, because only they know what counts as correct. That set is the acceptance test: not «it looks clever», but an agreed, measurable share of right answers.

Then we track three numbers: how many answers carry a citation, how often the system said «I don't know» and handed over, and how often it was wrong. Every model or prompt change is replayed against that set, or an «improvement» quietly breaks yesterday's work. For the first weeks it runs in shadow mode, answering to a log while a human answers the customer.

Our advantage here is engineering, not marketing: we install the CCTV and access control ourselves, so camera and model are designed together. Reliable number-plate reading needs roughly 100–150 pixels across the plate width — that one figure fixes your focal length, standoff distance and mounting angle. No network recovers detail the sensor never captured.

We take the stream over RTSP/ONVIF and decide where inference runs: a light detector on an edge device beside the camera keeps latency low and video off the network; heavier work goes to a GPU server. The output lands where it is needed: a barrier command, an operator alarm, a stock record.

Uzbek personal data law requires personal data of Uzbek citizens to be processed in databases physically located in the country. So we separate the zones at the outset: what may go to an external API, and what never leaves the perimeter.

The trade-off is honest: cloud models are strongest at Russian and Uzbek, but your data travels. Open weights — Qwen, Llama, Gemma — run on your hardware: a 7–8B model quantised fits in 8–16 GB of VRAM, a 70B one needs upwards of 40 GB. Weaker in Uzbek, heavier on capital cost, but nothing leaves the building.

Schedule is governed by the state of your data far more than by the choice of model. If procedures live in one unstructured PDF and sales history sits in three exports that will not reconcile, the bulk of the work is not machine learning but making that usable. We say so during scoping, not after signature.

Second, what the system may do: answer only, or also raise a CRM deal, reserve stock, issue an invoice. Every action is an integration and its own cost of being wrong. Third, rarely mentioned: AI has a standing cost — token bills or GPU depreciation, monthly. We show it before launch.