Ground truth before heuristics
We build a gold set first, with a held-out part scored once per milestone. Nothing lands in a pipeline layer without a before-and-after row against it.
Applied AI · Computer vision
Applied AI that turns cameras into measurements.
TKO Ventures builds AI for the physical world. We take footage you already have and turn it into structured, measured data: who was where, when, and what it means. Every system we ship is evaluated against ground truth.
Loading possessions…
Simulated tracking data
The replay could not load its data.
A 2D court view of the kind our basketball pipeline produces from broadcast video. Dashed lines join each attacker to his nearest defender, and the green ring marks the most open attacker. These possessions are simulated.Explore all eight.
What we build
From pixels to decisions, end to end.
Most of the value in a camera feed is lost between the model and the person who needs an answer. We build the whole path: data, models, evaluation, and the screen someone actually uses.
- 01Perceive
See what the camera sees
Detection, segmentation, multi-object tracking, re-identification and reading text in the scene, on footage as it really is: moving cameras, cuts, occlusion, glare.
- 02Measure
Turn pixels into metres
Camera calibration and per-frame homography put every object on a real-world plane, with units, timestamps and identity, so the output can be counted, compared and audited.
- 03Decide
Answer the question that was asked
Positions become the insight a person acts on: who was open, where space opened up, what changed between two periods. Each claim links back to the frames behind it.
The vehicle
A small studio for applied AI ventures.
TKO Ventures is how we take an applied AI idea from a hunch to a measured system. We pick problems where the evidence is already on camera and the people involved have a real decision to make: a coach preparing for next week, an operator planning a shift, an engineer looking for the cause of a failure.
We work with the people who own the problem. We build the full system in the open, with its gold set, metrics and costs kept alongside the code, and we grow the projects that hold up into products. Senior engineering and AI agents work side by side, which keeps the team small, the loop fast and the paperwork honest.
- ModelVenture studio: build, measure, then spin out or partner
- FocusComputer vision and spatial data from real-world footage
- Working styleGold set first, every number reproducible from a config
Approach
How we build things that hold up.
Six rules we follow on every project. Each one exists because ignoring it once cost us time.
Every layer is swappable
Each stage is a backend behind one interface and one config. We compare a new detector, tracker or reader by its measured score, not by our opinion of it.
Open weights, licences checked
Production runs on open-weight models fine-tuned on the client's own data. We check every licence before the first download, and closed APIs stay out of the per-frame path.
Show the capability early
A working end-to-end demo in days, labelled honestly as automatic and uncorrected. Then the slower work of measuring and improving it, layer by layer.
Human time is the budget
Agents do the first pass on labelling and review. People correct only what is flagged, in sessions sized to their week. Each correction goes back into the training data.
Compute where it is cheapest
We run inference locally where we can and use per-second GPU rental for training and CUDA-only stages. We log the cost of every job, so the bill is never a surprise.
Case study
Reading a basketball game from the TV camera.
An offline pipeline that takes broadcast footage of Polish top-league basketball and puts every player on a 28 × 15 m court, with team, identity and jersey number. It started from public models. The court model is now being fine-tuned hall by hall, and every step is scored against a gold set with a held-out game.
Read the case study- Held-out HOTA, baseline
- 0.648
- Track fragments after SAM 2.1
- 752 → 190
- Re-identification across cuts
- 87%
- GPU spend for the gold-set prefill
- $3.60
Where it applies
Same building blocks, other floors.
Detect, track, map to a plane, explain. The toolkit behind the basketball work carries over to any place where a fixed or moving camera watches people or objects on a known surface.
Sport
Positions, spacing, coverage and workload from broadcast or venue cameras.
Venues and retail
Flow, dwell and queueing measured on the floor plan rather than estimated.
Industrial sites
Zones, near-misses and throughput on lines and yards that already have cameras.
Mobility
Counts, trajectories and conflicts at junctions, platforms and car parks.
Contact
Have footage and a question it should answer?
Tell us what the camera sees and what decision you need to make. We will tell you honestly whether today's models can get you there and what a first measured version would take.