AI engineering · Tallinn, Estonia

Most AI projects die as demos. Let's scope one that doesn't.

Describe a process that eats your team's week. The scoping agent below reads it the way I would on a first call — signals, architecture, the thing most likely to kill it — and gives you an honest answer, including "this doesn't need AI."

usturai · scoping-agent local
>
runs in your browser · nothing is sent anywhere
    Why this exists

    The gap is never the model

    You have probably run an AI experiment already. Something that impressed people in a meeting, then never touched real data.

    What breaks a project is everything around the model: integration with the systems you actually depend on, the inputs nobody anticipated, keeping a human in the loop where the stakes demand one, and knowing what it costs per run at real volume. Closing that gap is the entire job — and it is a delivery problem before it is a machine-learning one.

    The agent above is a small demonstration of the same instinct: it does the cheap pass first, commits expensive effort only where the cheap pass found something, and tells you when the answer is no.

    Services

    Three ways to work together

    Build

    AI agents & automation

    Custom agents for document processing, research, support and internal workflows — RAG architectures, LangChain, the Anthropic API — wired into your existing stack rather than bolted beside it.

    Advise

    AI systems consulting

    An independent read on where AI actually pays off, a cost and ROI model with real numbers in it, and a delivery plan — before anyone writes code. Useful if you have been sold AI for its own sake before.

    Scale

    Meta-agent & orchestration

    Agent systems that build, test and deploy other agents, with cost tracking and profitability modelling designed in from the first commit rather than discovered in the invoice.

    Method

    How the work runs

    01

    Discover

    A working session on your real process — not a discovery questionnaire. Output: a map of where AI removes friction and where it would just add a moving part.

    02

    Prototype

    A working version in weeks, tested against your real data and its real edge cases. If it fails here, you lost weeks instead of quarters.

    03

    Deploy

    Production hardening, monitoring, integration and handover — including cost per run, so the finance conversation happens before rollout rather than after.

    04

    Iterate

    Systems drift as your data and users change. Ongoing refinement, or a clean handover to your team — whichever you actually want.

    Evidence

    Built, not slidewared

    Case · Adventure Highlight Agent

    Finding the twelve good seconds in sixty gigabytes of footage

    An end-to-end AI video pipeline that scans hours of raw footage and surfaces the moments worth keeping. It runs a fast, cheap vision pass to find candidates, then an expensive precision pass only where the first pass hit — the architectural decision that makes the economics work at all. The same two-pass instinct drives the agent at the top of this page.

    Built solo end to end: architecture, backend, review interface, and a cost model that recalibrates itself against real runs instead of guessing. Human-in-the-loop review throughout, because full automation would have been the wrong answer here.

    60 GB+
    per project
    2-pass
    inference design
    ± 15%
    cost estimate drift
    1
    engineer, end to end
    Node.jsClaude Vision APIffmpegWhisperSSETauri
    Who

    Who you would be working with

    UsturAI is the AI practice of a Tallinn engineer with two decades of running IT for organisations that cannot afford downtime — including an ERP implementation in recent years — now working full time in applied AI engineering.

    That combination is the point. Building an agent is increasingly common. Knowing how a real organisation absorbs a new system — who resists it, what breaks in month three, which integration nobody mentioned at kickoff — is what decides whether it ships. You get the person who does the work, not an account manager relaying to a delivery team.

    Working stack: Node.js, Python, RAG architectures, LangChain, FAISS, the Anthropic API, PyTorch — and the judgement to say when none of them is the answer.

    Privacy

    What this page does with what you type

    Ran the assessment? Let's pressure-test it.

    No deck, no discovery fee, no pitch. Thirty minutes on whether the thing is worth building — including an honest no.

    Start a conversation

    or write directly to hello [ät] usturai.com