AI business builder · Inference economics · Enterprise GTM

Jake Scherbak

I make AI pay off for banks

Co-founder of a UK AI research lab. I build AI-native businesses, cut the cost of every successful AI task and take enterprise AI from pilot to production.

See my AI work ↓ Get in touch
$50M+in closed B2B deals
120people in sales teams I led
8layers in our AI inference stack
13companies I built or advised
The network on the right runs a live forward pass. Move your cursor over it to fire neurons.
Whooo?

Top management roles for the past 11 years.

Chief Business Development Product Revenue Doesn’t Matter Officer at several companies.

Near 3000 years ago, in a previous life, I ran complex B2B IT sales. These days I co-found, partner and advise.

Now · Co-founder

Making every successful AI task cheaper.

Cognia AI Lab

🇬🇧 A UK research lab working on AI inference optimization.

  • North-star metric: cost per successful outcome, not cost per request.
  • 8-layer optimization stack: from GPU infrastructure to ops governance.
  • Research collaborations with university scientists in the UK (Cambridge, LIMS, NE), MENA (MBZUAI) and China (Peking).
  • Inference Optimization: the lab’s direction that makes every successful AI task cheaper.
Inference Optimization ↗ cognia-lab.com ↗
The 8-layer stack
request trace
    AI · What I study

    The real price of AI is the price of a solved task.

    01 · Research

    How AI speeds up scientific discovery

    • How to turn a researcher’s thinking into technology: break it into processes that machines can reproduce.
    • A catalogue of 173 AI tools for science across seven research stages, from AlphaFold 3 in biology to Lean and DeepSeek-Prover in mathematics.
    • How to make every business task cheaper, opening up scenarios that did not pay off yesterday.
    173 tools · 7 research stages
    02 · Economics

    The economics of inference

    • The core question: what does a successfully solved task cost across the whole AI vertical, from infrastructure to result? The price of a single model call matters less.
    • The trap we described: the price per request falls, and the share of solved tasks falls with it. Savings on paper, losses in practice.
    cost per request cost per solved task pricier model cheapest model
    Cost per request×1.0
    Cost per solved task×1.0
    Illustrative. The false-economy pattern.
    What we achieved
    8layers of inference optimization, from GPU infrastructure to ops governance
    79components
    237+production-proven techniques
    ~600metrics
    We built the product
    • North-star metric: Cost per Successful Outcome.
    • Diagnostics across five loops:
    Serving & infrastructureWasted demandRetrieval & toolingToken economicsQuality leakage
    • How we work: a one-week audit, then an 8-week sprint to the first measurable result.
    Audit · 1 weekSprint · 8 weeksMeasurable result
    We built a research network
    • Direct R&D contracts with LIMS and Northeastern University London.
    • Cambridge Service Alliance: research, publications, events. The first joint paper is on token economics.
    • In progress: Peking University and MBZUAI.
    Surf · Telegram channel

    My thoughts and notes on AI.

    Inference engines, model economics, agents and compute, one picture at a time.

    Read the channel on Telegram ↗
    Businesses

    Companies I built.

    How? · What I bring

    AI that pays off, sold and shipped.

    Four AI capabilities on top of four business ones. Each comes with a track record.

    AI
    Business
    What?

    People from these companies, surprisingly, do not refuse me.

    My biggest victory in life, besides my daughter, of course.

    I’m on the web

    Say hi.