We are in your building while we build it, sitting with the people who do the work. After go-live we stay: we monitor it, fix what breaks, and re-version it until it meets the targets we signed together, measured on your data. If one is the wrong target, we say so and re-cut it with you.

What working with us actually looks like

Every model we ship crosses the same five stations. The order matters: nothing is quantised before it is profiled, and nothing is promoted before a held-out evaluation says the compressed model still does the job.

Graph fusion, operator selection and engine build per device and precision: ONNX to TensorRT, or a compiled runtime where the target has no GPU at all.

Compile

Layer-level latency and memory on the actual target, under the actual input distribution, with the thermal state logged alongside.

Profile

Structured pruning of channels and heads, then knowledge distillation from the full model into the smaller one, so capacity is removed where the profile says it is idle.

Prune & distil

Post-training quantisation with representative calibration data first; quantisation-aware training only where INT8 or FP8 costs accuracy the evaluation will not forgive.

Quantise

The compressed model re-runs the full held-out evaluation, event-clustered and base-rate-adjusted; a regression anywhere sends it back a station.

Verify

The optimisation bench

Describe the signal, the device and the decision it has to support. We will tell you which of these methods apply, what a baseline would look like, and what we would want to measure first.

Bring the lab a problem

Speech and sound as signals first, then as language

Making a model fit the box it has to live in

Language models put to work on documents that never leave the building

Time series, recurrence and the models that replaced the ones we grew up on

Multi-object tracking where the frame budget is the specification

Depth, geometry and what a camera can be trusted to know

Questions the lab gets asked

Research into what deterministic control can and cannot do for small AI agents: bounded steps, closed vocabularies where honest abstention is a correct answer, and claims that need witnesses.

Deterministic runtime control

Open-weight models, quantized and served on commodity hardware, fine-tuned for industrial root-cause analysis so the plant's telemetry never leaves the plant.

Local-model inference stack

A data engine that turns real fault-labelled wind and solar telemetry archives and physics-grounded synthetic plants into training and evaluation worlds whose answers are checkable.

Training worlds with verified ground truth

An edge co-processor line for high-throughput vector retrieval and real-time drift and anomaly monitoring, designed to sit beside a standard industrial PC.

Edge acceleration programme

On-premise industrial AI, from silicon to answer

Each section below is a working vocabulary: the methods we reach for, what they are for, and where they are running today. Nothing here is a claim about accuracy. Every number lives in the customer's evaluation suite, not on a marketing page.

Six disciplines, one bench

Every experiment includes the boring baseline: seasonal-naive, a linear model, last year's architecture. A new method ships only when it beats the baseline on held-out data across seeds, reported as a mean with its spread.

Baselines before novelty

Thresholds, metrics and the definition of a pass are written down before the run. A result that needs the rules changed afterwards is a finding about the rules, not a success.

Pre-registered evaluation

Simulated scenes, rendered sensors and generated edge cases fill the gaps real data cannot, and every synthetic result is confirmed against a real capture before it counts.

Synthetic where reality is expensive

Weights, training code, evaluation harnesses and traces are delivered into the customer's repositories and hardware. We keep the method; you keep everything it produced.

Customer-owned artefacts

How the lab works

The engagement mechanics: discovery, build, rollout, monitoring, handover.

How we work

The platform

Where this goes next

We will say whether a programme is warranted, or whether a bounded system is enough.

Tell us where the work jams.

What changes, and what does not

Questions owners ask

We first observe the real workflow. If software is not the bottleneck, we say so and stop the programme before building.

Solve the right problem first

Systems propose; people decide. Cost-bearing actions wait for a named approver recorded before go-live, while the team can still challenge the rule.

A person approves anything consequential

The system runs inside your existing servers, tenancy or site. That starting constraint decides the architecture and carries no premium.

Runs where the data lives

Each engagement names its target categories, the measurement method and the review points, and they are measured against your own baseline. We do not carry a number from another client onto your contract, and we do not publish one.

Signed targets, measured on your data

Everything is versioned and documented for handover. A retainer remains your choice when the next problem is worth continuing together.

Handover is the goal, not the retainer

Principles we hold to

Write to us about a system you built, what broke after launch, and what you did about it.

How to apply

Introduce yourself anyway. Tell us about a system you built and kept running; the roles above are where we are growing, but strong builders make their own case.

None of these fit?

You build the agentic systems we sell: deterministic control engines, on-prem model serving, and agents that cite their evidence before anything consequential happens.

AI Engineer

You turn plant telemetry, camera streams and operational records into honest evidence: chance rates beside every accuracy, and no figure a client cannot trace.

Data Scientist

You ship our systems into customers' buildings and stay until they run: servers, cameras, edge boxes and the people around them. Client sites out of London and İstanbul.

Forward Deployed Engineer

You work in the lab behind the products: fine-tuning local models against our control engines, building training worlds with checkable answers, and putting negative results on the record.

ML Research Engineer

Open positions

Not a prototype that demoed well. Something that survived real users, real load and the month after launch, with you still responsible for it.

You have shipped and kept something running

The person doing the work knows where it breaks. If you can spend a morning on a floor without steering the conversation to your solution, you will do well here.

You can sit with an operations owner and listen

Decisions, trade-offs and the reason a thing is the way it is. Handover is our product; undocumented cleverness is a liability we would have to unpick later.

You write things down

If a figure cannot be traced to a source we can show a client, it does not ship. We would rather lose a sentence of marketing than carry a claim we cannot stand behind.

You say no to fabricated numbers

What we look for

London and İstanbul, with time on client sites. Remote for the parts of the work that do not need a factory floor.

Where we work

Privacy settings

A record holds the time, the site language, the notice version, the page you were on, a coarse description of your browser and system, and the purposes you allowed. It does not hold your IP address, in the clear or hashed.

Consent, and the record we keep of it

Would count page views and report which pages are read, so we can see what is useful. Nothing loads until you allow it.

Analytics

Uses the details you send in a contact or proposal form so a person can read your message and reply to it. Asked for at the form itself, when you send it.

Contact requests

Keeps your session working, keeps the site secure and stores the record of the privacy choice you made. Runs on every visit and cannot be switched off.

Essential

The page you were looking for is not here. It may have moved, or the link may be wrong. The three routes below cover most of what people arrive looking for.

Name

Tell us about your project and we'll get back to you within 24 hours.

Request Sent!

Get a Quote

Koca Ventures builds software and AI systems, and provides fractional CTO and CFO services.

Work here

We want high-ownership builders and operators who care more about keeping real systems useful than titles or theatre. We work on site, write down decisions, and use a paid, scoped problem instead of a long take-home.