The studio

Fourteen people.
Zero demos that die.

Aether is an applied-AI studio, not a research lab and not a body shop. We take on eight to ten engagements a year, staff them with people who stay from first prototype to post-launch, and measure ourselves on what kept running after we left.

Founded
2021
Team size
0
Systems in production
0
Engagements per year
8 – 10

Why we
exist

Aether started because two of us kept getting the same phone call: a team had built an impressive model prototype, demoed it to leadership, won a budget — and then discovered nobody could tell whether it was actually safe to put in front of customers.

We spent our first two years inside other companies' infrastructure, learning where AI systems actually break: retrieval that quietly drifts as documents change, evaluations that measure the wrong thing, prompt chains with no owner, and the hundred unglamorous integration details that decide whether a model ever reaches a real user.

So we built the studio we wished had existed. Small enough that the people who scope your project are the people who build it. Senior enough that nobody learns on your budget. And stubborn enough to say no when a project is a demo dressed up as a product.

We do not sell models. We sell the boring, durable work of making a model useful to a business — and then handing the keys over.

What we hold to

Four operating principles. They are not values on a wall — they are the reasons we have turned work down and the reasons clients keep calling back.

  1. Ship the smallest honest thing

    We build the narrowest system that can prove or disprove the business case, put it in front of real users, and measure. Expansion is a reward for evidence, not a default.

  2. Evaluation before elegance

    If we cannot measure whether the system is right, we do not ship it. Eval sets, golden data and regression checks get built in the first sprint, not bolted on after launch.

  3. Your engineers, not ours, at the end

    Every engagement includes handover: documentation, runbooks and pairing sessions so your team owns the system. We would rather be unnecessary than indispensable.

  4. Say the uncomfortable thing early

    If a use case will not work, if the data is not there, or if a cheaper non-AI solution wins, we say so in week one — not after the invoice.

The people you'd actually work with

No account managers, no offshore handoff, no rotating cast. These are the leads on every engagement; the wider bench fills in the specialisms a project needs.

  • Iris Okonjo

    Founding partner · Applied ML

    Spent six years building ranking and retrieval systems at a marketplace at scale before deciding that the interesting problems were in messier, smaller companies. Leads evaluation design on every engagement.

    • Retrieval
    • Evaluations
    • Ranking

  • Marek Dolinský

    Founding partner · Platform

    Infrastructure engineer by training, sceptic by temperament. Designs the deployment, observability and cost-control layer that keeps agentic systems from becoming unbillable science projects.

    • Platform
    • Observability
    • Cost control

  • Priya Raman

    Lead · Product & research

    Turns ambiguous business problems into testable AI scope. Previously a product lead in healthcare ops, which is where she learned that a model is only as good as the workflow it sits inside.

    • Discovery
    • Workflow design
    • Research

  • Tomás Ferreira

    Lead · Agentic systems

    Builds tool-using agents and the guardrails around them. Keeps a running catalogue of agent failure modes that is, frankly, more useful than most papers on the subject.

    • Agents
    • Tooling
    • Guardrails

  • Ana Gutierrez

    Lead · Data engineering

    Owns the pipeline layer: ingestion, chunking, lineage and the retraining loops that make a knowledge system stay accurate six months after launch. Allergic to undocumented transformations.

    • Pipelines
    • Lineage
    • Data quality

  • Dele Bankole

    Lead · Safety & governance

    Runs red-teaming, policy review and the human-in-the-loop patterns that keep regulated clients out of trouble. Former risk analyst, which explains the uncomfortable questions.

    • Red-teaming
    • Governance
    • HITL design

A deliberate shape

We are not trying to become an agency of two hundred. The studio is structured so the same six leads stay close to the work, and the wider bench only grows when a specific problem demands it.

Every engagement runs as a pod: one product lead, one applied-ML engineer, one platform engineer, plus the specialists the problem needs. Pods have one client at a time. Nobody is spread across four projects pretending to be everywhere.

That constraint is the point. It is also why we cap intake at eight to ten engagements a year — and why, when we are full, we tell you honestly and put you in touch with people we trust.

      • iris-okonjo
      • marek-dolinsky
      • priya-raman
      • tomas-ferreira
      • ana-gutierrez
      • dele-bankole
      • retrieval-and-support
      • agentic-operations
      • decision-modelling
      • managed-ai-ops

Six leads across four pods. Specialists join from a bench of eight.

We have capacity for
one more engagement this quarter.

If you have a problem that genuinely needs a model — and the discipline to measure whether it worked — we would like to hear about it.

Contact details