Data
Every deployment produces more movement data, and every new hour of data makes the model harder to replicate.
Investors
Movement is the last major data modality without a foundation model.
Text has GPT. Images have CLIP. Proteins have AlphaFold. WithoutBall is building the intelligence layer for human movement — self-supervised, validated on real data, and bootstrapped to a working proof of concept.
Thesis
Every data modality gets its foundation model. Movement is next.
One self-supervised model, trained on skeletal motion, that learns what skilled, healthy, and deteriorating movement looks like — and transfers to injury signals, scouting, and consumer movement health without task-specific labels.
| Modality | Foundation model | Status |
|---|---|---|
| Text | GPT | Established |
| Images | CLIP | Established |
| Proteins | AlphaFold | Established |
| Movement | WithoutBall — Hierarchical Motion Transformer | Open — being built now |
Why now
Five independent trends, one conclusion.
None of these trends is ours to control. Each matured on its own; together they make a movement foundation model buildable now — and they make the window a first-mover question.
Pose estimation matured
Extracting reliable skeletons from ordinary video was a research problem for decades. Since roughly 2022 it is production reality: modern estimators read 2D skeletons from any camera at real-time rates. The raw input a movement model needs became a commodity.
The foundation-model paradigm is proven
Self-supervised models trained on large unlabeled corpora now outperform task-specific systems in language, vision, speech, and protein structure. The recipe is established; applying it to a new modality is no longer a leap of faith.
Movement video became usable at scale
Movement video exists in abundance across the sporting and everyday world. Once it can be converted into clean skeleton data, that abundance becomes the raw material of a foundation model. Three years ago it was unusable; today it is not.
Sports-technology spending is accelerating
Professional clubs already buy performance technology at scale, and analytics budgets grow roughly 25% a year. The demand for analysis beyond volume metrics is explicit — analysts know that load numbers alone cannot explain how a player moves.
Movement health demand is demographic
1.4 billion people will be over 60 by 2030 (WHO). Musculoskeletal disorders are already the leading cause of disability worldwide, costing European health systems over €200 billion a year. The need to measure and maintain movement quality is structural, not cyclical.
These trends are independent — none depends on us, and none reverses easily. Our read: the window for establishing the movement foundation model is open now, and it narrows as the recipe becomes obvious to more research groups.
Proof of concept
Validated, not projected.
One training campaign answered the core scientific question — can a transformer learn the structure of human movement from raw skeletons alone? These numbers describe what already happened, not a plan.
Built8×NVIDIA H100 GPUs in one continuous run
Built14days of uninterrupted distributed training
Built~2,700GPU-hours of compute
Built0labels — fully self-supervised
Built3capabilities validated on held-out real data
Built4co-founders — research, engineering, elite football
These are observed outputs of the proof-of-concept model on held-out data — not projections. The first published milestone on the roadmap is a head-to-head benchmark against the strongest available off-the-shelf pipeline.
De-risked
What the proof of concept already answered.
Deep-tech risk is a queue of open questions. These five are closed.
| Risk | The question going in | Where it stands |
|---|---|---|
| Convergence | Does the architecture converge on real, noisy skeletal data? | Yes — training remained stable across the full 14-day campaign. |
| Objective | Can self-supervision learn movement structure without labels? | Yes — self-supervised training on raw motion produced embeddings that encode speed, stability, and joint asymmetries, with zero labels involved. |
| Missing data | Can one model handle occlusion and gaps natively? | Yes — the model restores occluded and missing joints natively; robustness to gaps is built in, not patched on. |
| Execution | Can this team run multi-week distributed training at scale? | Yes — the founding team planned, executed, and monitored the 8-GPU campaign end to end. |
| 2D→3D | Does lifting ordinary video into 3D fit the same architecture? | Yes — the same model restores the depth a camera discards: 2D video in, consistent 3D skeletons out, in one forward pass. |
Market
Two markets, one model.
Movement is the integrated output of the musculoskeletal and nervous systems — one of the richest health and performance signals there is, and one of the least used. The tools clubs and consumers buy today measure volume: distance, load, steps, heart rate. The question none of them answers — how is this person moving, and what does it mean — is the market. One R&D investment feeds both sides of it.
Professional football — B2B
| Signal | Figure | Source |
|---|---|---|
| Sports analytics | $4.5B (2025) → $14B+ (2030), ~25% CAGR | Grand View Research; MarketsandMarkets |
| Injury burden | ≈50 injuries per squad per season at the top European level — hundreds of player-days lost | UEFA Elite Club Injury Study |
| Injury economics | €3M–€10M — the full cost of one major injury to a top club | Industry estimates: medical costs, wages in recovery, transfer-value impact |
| Performance-technology spend | Millions of euros per club per year, growing ~25% annually | Published club technology budgets |
Personal movement health — B2C
| Signal | Figure | Source |
|---|---|---|
| Fitness technology | $12B+ (2025), growing ~12% annually | Statista |
| Wearables | $60B+ | IDC |
| Regular runners | 150M+ worldwide; annual injury incidence estimated at 37–79%, mostly biomechanical | World Athletics; IHRSA; systematic reviews (van Mechelen; van Gent et al., Br J Sports Med) |
| Gym members | ~200M worldwide performing loaded movements where form drives injury risk | IHRSA industry estimates |
| Gait analysis today | €500–€2,000 per lab session; accessible to under 1% of those who would benefit | Published laboratory session rates |
| Demographic tailwind | 1.4B people over 60 by 2030; musculoskeletal disorders already cost European health systems €200B+ a year | WHO |
The B2B wedge, counted in organizations
~98clubs in Europe's top five leagues
200+clubs in European second tiers
~50national-team programs
For a club, preventing a single major injury pays for years of analytics. For a runner, a phone camera replaces a lab visit that fewer than one in a hundred athletes can access today.
Platform
One model, a platform of products.
Every product below is a lightweight head on the same movement embeddings. Each additional product reuses the shared representation instead of a new stack — the marginal product gets cheaper to build.
Club analytics — B2B
- InjuryRadar
- SkeletonAPI
- MoveSearch
- MoveID
Consumer movement health — B2C
- Injury prevention
- Running biomechanics
- Rehabilitation monitoring
- Movement readiness
384-dimensional movement embeddings — one shared representation
Hierarchical Motion TransformerSelf-supervised foundation model · zero labels
- InjuryRadarLead productRoadmap
Biomechanical early warning for the whole squad — subtle drift surfaced weeks before it becomes an absence.
- SkeletonAPIRoadmap
Clean 3D movement data from ordinary video, in a single forward pass.
- MoveSearchRoadmap
Semantic search across a season of movement — query behavior, not tags.
- MoveIDRoadmap
Player identification from gait alone.
- Injury preventionRoadmap
Personal biomechanical early warning from a phone camera.
- Running biomechanicsRoadmap
Form analysis and movement-quality indicators for every run.
- Rehabilitation monitoringRoadmap
Objective recovery tracking between clinic visits — never a medical diagnosis.
- Movement readinessRoadmap
A daily readiness signal built from how you actually move.
Landscape
Five busy categories, one empty layer.
Each existing category solves a real problem and has built a real business. Each also stops at the same line: it captures movement without understanding it. Notice the pattern in the last column — it is the position we occupy.
| Category | Representative players | What they do | What they cannot do |
|---|---|---|---|
| Tracking systems | Catapult; Stats Perform; Second Spectrum | Position, speed, distance, load, event detection | Read movement quality, or see how a player's mechanics change |
| Video analytics | Hudl; Wyscout; InStat | Tag events, retrieve clips, build statistical dashboards | Search by movement similarity or detect biomechanical patterns |
| Clinical gait labs | Vicon; Qualisys | Gold-standard 3D capture, force plates, expert reads | Scale beyond the lab — six-figure installations, a handful of subjects a day |
| Fitness wearables | Whoop; Garmin; Apple Watch; Oura | Steps, heart rate, sleep, recovery scores | Assess how you move rather than how much you moved |
| Pose-estimation libraries | Open-source pose estimators | Detect joint positions in video, fast and free | Understand what the detected pose means — detection without understanding |
| Movement intelligence | WithoutBall | Learns the structure and meaning of movement itself | — the layer every row above is missing |
WithoutBall sits on top of pose estimation and beneath applications — a complement to tracking and video suppliers, not a competitor. The moat is not one barrier but four loops that tighten with every month of operation.
Moat
One position, four compounding loops.
Tracking systems measure. Video tools tag. Labs analyze. Wearables count. Pose estimators detect. Nobody understands.
Products
New products are lightweight heads on shared embeddings — each launch widens the surface a rival must match.
Users
Longitudinal per-person baselines grow more valuable with use — switching away means abandoning your own history.
Trust
Clubs and clinicians extend trust slowly and keep it long. Evidence-first answers compound it month after month.
The loops explain why the moat deepens. Four barriers explain why it exists at all — why an incumbent cannot simply assign a team and catch up:
Research-grade architecture
The Hierarchical Motion Transformer is a purpose-built architecture, not an off-the-shelf model pointed at new data. Sports-analytics incumbents are software organizations, not transformer research labs — standing one up is a multi-year transformation, not a product sprint.
Self-supervision for skeletons
Making self-supervised training work on skeletons in motion — so the model learns biomechanics rather than shortcuts — is an open research problem, not a documented recipe. Solving it took representation-learning depth guided by one of Europe's most cited ML scientists.
Biomechanical domain knowledge
A model can be technically correct and biomechanically meaningless. Knowing what the model must be sensitive to and what it must ignore requires elite-football depth — every architecture decision here was validated against real coaching and medical workflows from day one.
Compute and training infrastructure
Foundation models demand sustained GPU-cluster access and the distributed-training engineering to use it reliably for weeks. We have run a ~2,700-GPU-hour campaign end to end — infrastructure most sports-technology companies have never operated.
The replication test: a rival needs the research team, the compute, a pipeline that converts movement video into training-grade skeletons at scale, relationships across professional football to keep the model honest, and the patience to fund a year of R&D before revenue. Each piece is attainable. The combination, in one organization, is the moat — and it tightens as data, products, and trust compound.
Data strategy
Independent of any single partner, by design.
A foundation model is only as good as its data strategy. Ours is built on one principle: no single partner, league, or provider may constrain the model. Three channels, each with a different job.
| Channel | What it provides | Status |
|---|---|---|
| Movement data at scale | A massive and growing corpus of movement video, converted into skeleton data — diverse across sports, body types, ages, and abilities. Diversity is what a foundation model needs to generalize. | Scaling now |
| Professional collaborations | Depth and rigor: relationships across professional football that keep the model grounded in real coaching and medical workflows. | In active discussions |
| Consented user data — future | Once products ship, deployments can — with explicit consent — contribute movement data back to training, on GDPR-compliant foundations designed in from day one. | Designed, not yet live |
More usersMore movement dataA better modelBetter products
…which attract more users
Data-source specifics are confidential as a matter of policy, for the same reason the strategy works: they are part of the moat. Detail is available to qualified investors in the data room, under NDA.
Team
Four founders, one intersection.
Replicating WithoutBall requires research-grade machine learning, elite football domain knowledge, and production engineering in one room. Each is scarce. The intersection is the moat.
CEO · Co-Founder
Kemal İnecik
Machine learning PhD researcher at Helmholtz Munich & TUM, building deep generative models in one of the world’s leading AI-for-science labs. Author of more than ten machine-learning papers — in Nature, Cell, ICML, and NeurIPS among others, with more on the way. Designed the Hierarchical Motion Transformer from scratch and trained it end to end.
Co-Founder · ML & AI
Fabian J. Theis
Professor at TUM and Director at Helmholtz Munich. Leibniz Prize laureate — Germany’s most prestigious research award — ERC Advanced Grant holder, and one of Europe’s most cited scientists in machine learning for the life sciences. His group pioneered the application of foundation models to biological data: the closest methodological precedent to a foundation model for movement.
Co-Founder · Football
Şeref Çiçek
UEFA Pro licensed coach with twenty-two years at elite level. Led Beşiktaş to back-to-back Süper Lig championships and coached the Türkiye National Team. The domain expert who keeps the AI honest — a co-founder, not an advisor.
Co-Founder · Engineering
Özgür Ak
Mechanical engineer bridging hardware sensing, biomechanics, and scalable software. Builds the production line from raw movement data to served predictions — ingestion, distributed training, inference — grounded in the physics of movement, not just software abstractions.
Remove any one founder and the company loses a capability the others do not cover: the architecture, the scientific credibility and institutional access, the domain validation and market access, or the production systems. The complementarity is deliberate — and it is the hardest thing here to replicate.
Funding status
Bootstrapped to a validated foundation model.
100% founder-owned, no external capital, no debt. A seed round is in progress to fund model scaling, first club pilots, and the consumer beta.
| Position | Status |
|---|---|
| Ownership | 100% founder-owned — a clean cap table |
| External capital | None — bootstrapped through the proof of concept |
| Debt | None |
| Stage | Pre-revenue · deep-tech R&D |
| Round | Raising a seed round |
| Partnerships | In active discussions with professional football organizations; active research collaboration with Helmholtz Munich & TUM |
The large majority of new capital goes to R&D and engineering — our most effective sales tool is a working model demonstrated to the right person. Only a small share goes to marketing.
Roadmap
Milestones, not dates.
We sequence the plan to resolve the highest-uncertainty questions first, with the cheapest possible experiments. Phases advance on milestones, not on the calendar.
- 01
Validate Roadmap
We scale the model and the data pipeline, benchmark head-to-head against the strongest off-the-shelf alternative, and open the first pilot conversations.
Exit: published benchmark · 1–2 pilot clubs engaged
- 02
First revenue Roadmap
We convert pilots into paying deployments and iterate the product on real club feedback — testing willingness to pay, not willingness to talk.
Exit: 1–3 paying club deployments
- 03
Consumer beta Roadmap
With the technology and the B2B market both validated, we open a focused consumer beta — running-gait analysis first — while the club base grows.
Exit: consumer beta live · 4–8 club customers
- 04
Growth Roadmap
We scale both channels, open the platform API, and explore adjacent verticals beyond football.
Exit: compounding club and subscriber base
The order is deliberate. Phase 1 answers the hardest technical question — does the model beat the best off-the-shelf alternative? — with a cheap, binary experiment. Phase 2 answers the hardest commercial question: will clubs pay? Phase 3 commits to consumers only after both answers are in. If any phase fails, we learn early and cheaply — and the seed round is sized to fund the first two phases completely and begin the third, so there are multiple decision points before the capital is spent.
Metrics
What we measure — and what we refuse to project.
Pre-revenue deep tech is not measured in financial metrics. It is measured by whether the technology works, whether the team executes, and whether the market engages. We publish what has already happened; projections belong in the data room.
Today — technology and execution
| Metric | Status |
|---|---|
| Foundation model | Designed, built, and validated on real data |
| Proof-of-concept training | 8× NVIDIA H100 · 14 days · ~2,700 GPU-hours |
| Validated capabilities | Representation learning · skeleton reconstruction · 2D→3D lifting |
| Data pipeline | Scaling — movement video converted into training-grade skeleton data |
| Team | Four co-founders: ML research, elite football, engineering |
| Partnerships | Multiple professional-football discussions active |
| Public presence | withoutball.com live in three languages; early-access waitlist open |
Next — the proof points that matter
| Proof point | The question it answers | Sequence |
|---|---|---|
| Head-to-head benchmark | Does the model beat the strongest off-the-shelf pipeline on football-relevant tasks? | Phase 01 |
| First club pilot | Will a professional club deploy the product and keep using it? | Phase 01 |
| First paying club | Will a pilot convert to revenue after the trial period? | Phase 02 |
| Consumer beta | Can consumers understand and act on biomechanical feedback from a phone? | Phase 03 |
Unit economics will be reported against public sports-technology and consumer-health comparables once there is revenue to report. Until then we publish no financial projections — a discipline we consider a feature, and the same honesty the status chips enforce across this site.
FAQ
The questions investors ask first.
Why won't a large incumbent simply build this?
Because the required combination is rare: a purpose-built research architecture, self-supervised training expertise for skeleton data, biomechanical domain depth, and multi-week distributed-training infrastructure. Sports-analytics incumbents are excellent software organizations — they are not transformer research labs, and standing one up is a multi-year transformation. The four barriers above are the long answer.
Where does the training data come from?
From movement data at scale — a massive and growing corpus of movement video, converted into skeleton data, deliberately assembled so that no single partner or provider constrains the model. Source specifics are confidential by design and are shared with qualified investors under NDA.
How is this different from pose estimation?
Pose estimation answers where the joints are, frame by frame. The foundation model answers what the movement means: it turns joint positions into representations that encode quality, style, symmetry, and change over time. Pose estimation is commodity input at the bottom of our stack; the learned representation is the asset.
What exists today, and what is roadmap?
The foundation model, its training pipeline, and the validated capabilities exist today — demonstrated on held-out real data. Every product built on top of them is labeled Roadmap, and every claim on this site carries an explicit status chip. We consider that honesty a sales asset, not a weakness.
Why professional football first?
It is the sharpest wedge: clubs already pay for performance technology, the injury-cost argument is immediate, and a co-founder spent twenty-two years at the top of the game — so we reach decision-makers directly. Human biomechanics are universal, which is why the same model then carries into consumer movement health.
What does the seed round fund?
Model scaling, the data pipeline, the first club pilots, and the groundwork for the consumer beta — with the large majority of capital going to R&D and engineering. The round is sized to fund the first two phases completely and begin the third. Terms and detailed financials live in the data room.
Data room
The detail lives behind an NDA.
The financial plan, cap table, partnership detail, data-source specifics, and technical benchmark results are available to qualified investors under NDA — info@withoutball.com.