AI agents & LLM platforms
Assistants that do work rather than demo it: tool use, retrieval over your documents, evaluations, guardrails and a cost you can predict.
AI software team
AI products that reach production.
A small senior team. We design, build and run AI-backed software — and the ordinary product engineering around it that turns a model into a business.
- 8+
- years in production ▸ FILL IN
- 20+
- products shipped ▸ FILL IN
- 2 weeks
- to a prototype you can click
- 1 day
- to a reply, and an NDA if you need one
What we do
Six directions
E-commerce & marketplaces
Catalogue, cart, payments and the back office that runs them — with search, recommendations and generated content where they pay for themselves.
Image & video generation
Diffusion wired into a product, not a notebook: fine-tunes, render queues, moderation, storage and a cost model that survives real traffic.
Computer vision
Detection, tracking, OCR and quality control on your own footage — trained on your data, deployed to your hardware, measured on a labelled set.
Dashboards & analytics
The read path first: events, aggregates and a warehouse designed before the screen. Then dashboards that answer in one screen, not five.
B2C apps & games
Consumer products where retention is the only honest metric: mobile apps, games, onboarding, and the analytics that say whether it worked.
How we work
Four steps, and you can stop after any of them
Every step ends with something you own and can take elsewhere.
01
Discovery
What the model must do, what data exists, what "working" means in numbers. You get a scope, an architecture sketch and a fixed quote.
3 – 5 days · fixed price
02
Prototype
The riskiest part first, running on a URL, with the evaluation set that says whether the quality is there before anyone commits.
2 – 3 weeks
03
Build
Two-week increments, each deployed behind a flag, with tests, monitoring and the numbers that show it did what it should.
6 – 12 weeks
04
Run or hand over
We keep it running — cost, latency, quality drift — or hand it to your team with docs and a fortnight of pairing.
monthly · cancel any time
Selected work
Shipped, and still running
Most of it sits inside somebody's company under an NDA, so it is described by shape and result rather than by client name. All cases in detail →
RUBTUR Portal — the concierge service’s internal portal
remove the manual assembly of documents for a request — quotes, letters to hotels and stay programmes, in the client’s own templates.
- 5
- document types from one request
- 4
- external integrations
AI agents for Beeline
take the routine off the teams — hand market monitoring, reporting, legal review and meeting minutes to agents.
- 8
- agent services in use
- 4
- document formats out
Agent platforms: Alfa-Bank and NanoForge
give the client a working agent platform rather than a prototype — and turn an agent engine into a product.
- 12
- agents and a builder in the demo
- 2
- platforms: web and desktop
Documents and primary records: Prodimex and Sotex
remove the manual entry of primary records and the manual preparation of legal documents — from a photo or a scan straight to a finished document.
- 9
- screens of the end-to-end process
- 4
- document types for the legal team
Mareven — AI Commerce & Intelligence
show what AI actually changes across a category team’s work — as something they can click, not as a deck.
- 8
- modules on one stand
- ~20
- live endpoints
Size data control
find the discrepancies between Lamoda’s size charts and the brands’ own sites — by hand that is 800 thousand rows of comparison.
- 150k
- SKUs under automatic comparison
- 800k
- rows instead of manual checking
The FashionJet → Yandex Market bridge
keep orders, prices and stock in sync between FashionJet and the marketplaces without manual exports.
- ~50
- operational tools
- 2
- marketplaces in one contour
Marketplace card audit: MP24.audit
show a brand where exactly its card loses to the competition — and what to fix first.
- 247
- SKUs in the Kuppersberg audit
- 6 / 9
- SKUs against competitors
B2B Video Studio
build advertising video scenes of a product with an “AI director”: one world for the whole clip, with prompts editable scene by scene.
- 1
- world across the whole clip
- 4
- stages from photo to video
RUBTUR Portal — the concierge service’s internal portal
remove the manual assembly of documents for a request — quotes, letters to hotels and stay programmes, in the client’s own templates.
- 5
- document types from one request
- 4
- external integrations
AI agents for Beeline
take the routine off the teams — hand market monitoring, reporting, legal review and meeting minutes to agents.
- 8
- agent services in use
- 4
- document formats out
Agent platforms: Alfa-Bank and NanoForge
give the client a working agent platform rather than a prototype — and turn an agent engine into a product.
- 12
- agents and a builder in the demo
- 2
- platforms: web and desktop
Documents and primary records: Prodimex and Sotex
remove the manual entry of primary records and the manual preparation of legal documents — from a photo or a scan straight to a finished document.
- 9
- screens of the end-to-end process
- 4
- document types for the legal team
Mareven — AI Commerce & Intelligence
show what AI actually changes across a category team’s work — as something they can click, not as a deck.
- 8
- modules on one stand
- ~20
- live endpoints
Size data control
find the discrepancies between Lamoda’s size charts and the brands’ own sites — by hand that is 800 thousand rows of comparison.
- 150k
- SKUs under automatic comparison
- 800k
- rows instead of manual checking
The FashionJet → Yandex Market bridge
keep orders, prices and stock in sync between FashionJet and the marketplaces without manual exports.
- ~50
- operational tools
- 2
- marketplaces in one contour
Marketplace card audit: MP24.audit
show a brand where exactly its card loses to the competition — and what to fix first.
- 247
- SKUs in the Kuppersberg audit
- 6 / 9
- SKUs against competitors
B2B Video Studio
build advertising video scenes of a product with an “AI director”: one world for the whole clip, with prompts editable scene by scene.
- 1
- world across the whole clip
- 4
- stages from photo to video
Content for Sportmaster and X5
produce card and campaign material at catalogue scale — batch food photography pipelines and img2video.
- 15+
- prompt iterations kept
- 4
- model families in the pipeline
Inspection of oil and gas infrastructure
inspect routes and facilities from a helicopter without processing terabytes of video by hand after the flight.
- a week → 1 h
- to process the data
- 3 → 1
- people in the process
Hard Collection Voice Analytics
find out whether collection operators follow the script — listening to every call by hand is impossible.
- 100%
- of contacts reviewed, not a sample
- 6
- script steps under control
agent-bi: text-to-analytics
give the business an answer from its own data straight away — without a request to an analyst and a wait in the queue.
- 4
- steps from file to takeaway
- 0
- lines of SQL from the user
A learning system for a large holding’s support department
bring three disconnected systems into one platform for learning and testing.
- 1500
- employees moved onto it
- up to 100%
- faster onboarding
An analytics system for an AdTech company
bring 20+ advertising platforms into one system — reports by team and profit by campaign.
- 30 → 500
- growth of the department’s headcount
- +300%
- department profit
ozhiv.ai — a meme video from the user’s photo
take the user from the first screen to a paid clip — the whole product.
- 11+
- clip templates in production
- 2
- landing languages: RU and EN
World of Heroes — a mobile RPG
build a hero-collector mobile RPG: MVP in six months, the release version in a year.
- 100 000
- installs across both stores
- 2
- regions on one backend
A social network for sports predictors
a public platform where an author publishes a prediction with their reasoning, and the result is visible to everyone.
- 20 000
- active users a day
- 400 rps
- at peak
Content for Sportmaster and X5
produce card and campaign material at catalogue scale — batch food photography pipelines and img2video.
- 15+
- prompt iterations kept
- 4
- model families in the pipeline
Inspection of oil and gas infrastructure
inspect routes and facilities from a helicopter without processing terabytes of video by hand after the flight.
- a week → 1 h
- to process the data
- 3 → 1
- people in the process
Hard Collection Voice Analytics
find out whether collection operators follow the script — listening to every call by hand is impossible.
- 100%
- of contacts reviewed, not a sample
- 6
- script steps under control
agent-bi: text-to-analytics
give the business an answer from its own data straight away — without a request to an analyst and a wait in the queue.
- 4
- steps from file to takeaway
- 0
- lines of SQL from the user
A learning system for a large holding’s support department
bring three disconnected systems into one platform for learning and testing.
- 1500
- employees moved onto it
- up to 100%
- faster onboarding
An analytics system for an AdTech company
bring 20+ advertising platforms into one system — reports by team and profit by campaign.
- 30 → 500
- growth of the department’s headcount
- +300%
- department profit
ozhiv.ai — a meme video from the user’s photo
take the user from the first screen to a paid clip — the whole product.
- 11+
- clip templates in production
- 2
- landing languages: RU and EN
World of Heroes — a mobile RPG
build a hero-collector mobile RPG: MVP in six months, the release version in a year.
- 100 000
- installs across both stores
- 2
- regions on one backend
A social network for sports predictors
a public platform where an author publishes a prediction with their reasoning, and the result is visible to everyone.
- 20 000
- active users a day
- 400 rps
- at peak
FAQ
The questions we always get
How do you charge?
Discovery is a fixed price. After that, a fixed scope for a fixed price or a monthly team rate — whichever fits the risk.
Who owns the code and the models?
You do, from the first commit — repositories, fine-tunes trained on your data, and infrastructure described as code so you can move it without us.
Our data cannot leave the country. Is that a problem?
No. We run open-weight models on your hardware or in your cloud region, and we tell you plainly what that costs in quality and in money.
What if AI is the wrong answer?
Then we say so in discovery and you have spent days instead of a quarter. A search index or a better data model often beats a language model.
Contact
Tell us what you are building
A paragraph is enough. If it is a fit we come back with questions and a first step; if it is not, we say so.