AI systems studio

We build the systems companies actually run on

Monitoring platforms, computer-vision products, reconciliation engines and on-chain protocols — designed, shipped and kept running in production.

Start a project See our work
8 Products shipped
7 Industries
2 Running in production today
About us

We are Polaim

A small engineering studio. We design AI systems, ship them, and keep them running in production.

Applied AI

Computer vision, LLM pipelines and retrieval, built into products people use every day.

Production first

Tests, migrations, monitoring and backups from the first commit. A demo is not the deliverable.

End to end

Research, architecture, build, deploy. One team, from the first sketch to the running system.

What we do

Our Solutions

What we build — grounded in what we have already shipped.

Computer vision

Recognition and measurement inside real products — from a photographed plate to a rendered 3D scene.

  • Image recognition
  • Measurement from a photo
  • Render pipelines
  • Shipped: photo-to-nutrition diary

LLM pipelines & agents

Extraction, classification and generation, with a person approving anything that leaves the system.

  • Structured extraction
  • Multi-provider routing
  • Human-in-the-loop
  • Shipped: AI tender studio

Monitoring platforms

Continuous ingestion, deduplication, scoring and alerting across noisy sources.

  • Semantic deduplication
  • Signal scoring
  • Alert routing
  • Shipped: cross-channel signal detection

Data & reconciliation

Deterministic engines where every number traces back to the file it came from.

  • Reproducible runs
  • Strict format validation
  • Full audit trail

Mobile products

Native and cross-platform apps taken all the way to the store, not to a prototype.

  • iOS and Android
  • Offline-first
  • Store releases

On-chain systems

Non-custodial protocols where the contract holds the logic and the keys stay with the user.

  • Smart contracts
  • Non-custodial by design
  • Integration test suites
How we work

A transparent process

Four steps. None of them exists to fill a slide.

01

Scope

We define the problem and what “working” means, in numbers, before anything gets built.

02

Architecture

The shape is decided before the code: data model, contracts, failure modes.

03

Build

Shipped in increments. Tests, migrations and CI from the first commit.

04

Run

Monitoring, backups, iteration. Staying alive in production is part of the work, not an afterthought.

Selected Work

Things we've built

Systems we designed and shipped — across media, consumer health, industrial B2B, fintech, travel, education and web3. Client and product names are withheld by agreement.

Media Intelligence

Signal detection across messaging channels

Editorial teams track hundreds of channels by hand — the important story drowns, the same news arrives in twenty reposts, half the feed is ads. We built a platform that looks for surges of attention rather than keywords: when independent sources raise the same topic at once, it becomes a ready news card with links to the originals, approved or rejected in one tap.

  • In production, multiple clients
  • Multi-tenant, DB per workspace
  • Python · FastAPI · PostgreSQL
Consumer Health

Photo-to-nutrition diary in a chat app

A photo of a plate becomes a diary entry in about ten seconds — the dish recognised, calories and macros counted, portion weight refined by computer vision using a coin in frame for scale. The system learns each user's typical portions, so accuracy grows with use, and every outgoing message passes an orchestrator that decides to send, delay or suppress it.

  • In production, paid subscriptions
  • ~120 modules · 170+ test suites
  • Node.js · Python CV · PostgreSQL
Industrial B2B

AI studio for tender proposals

Answering an exhibition-stand tender takes a manager, a designer and an estimator several days. We compressed it to hours: the brief is parsed into structured data, an AI design panel debates three concepts, the chosen one becomes real geometry in millimetres, gets rendered in 3D and auto-checked against venue rules and production limits. Nothing reaches the client until a human signs off.

  • Geometry is the source of truth, not the image
  • Venue rules enforced automatically
  • Next.js · TypeScript · Blender
Get in touch

Have a system to build?

Tell us what you need. We reply with an honest read on whether we are the right team for it.