Lead ai scientist - computer vision

IFREQ; Zaventem; Hybride; EUR 120 000 - 150 000; Plein temps

Résumé du poste

IFREQ in Brussels offers a Lead AI Scientist role focused on computer vision. You will own the CV R&D programme, define milestones, error budgets, and research direction, while leading the team to deliver models that run offline on mid-range Android devices.

You will shape the research roadmap, publish reliability and validity findings, and collaborate with the platform, backend, and product teams to ensure robust, protocol-bound scoring across drills.

Qualifications

• A research track record in computer vision with depth in video understanding and deployment.
• Edge deployment experience including on-device inference and latency budgets.
• Leadership of a small team and ability to drive a research programme from plan to production.

Responsabilités

• Own the computer-vision R&D programme: plan, milestones, error budgets, decisions.
• Decide what to model: temporal action localisation, pose estimation, tracking, per-rep events.
• Ship models for on-device inference and edge deployment; collaborate with platform and product teams.

Connaissances

Computer Vision

Leadership

Python

PyTorch

Video understanding

Formation

PhD in Computer Science

Outils

ONNX

TFLite

FastAPI

PostgreSQL

Description du poste

Location: Brussels, hybrid · Type: Full-time · Reports to: CTO · Team: you lead the AI track (growing under you)

The thesis

Lead AI Scientist — Computer Vision

Location: Brussels, hybrid · Type: Full-time · Reports to: CTO · Team: you lead the AI track (growing under you)

iFreq is building football's measurement standard: every number protocol-certified, video-proven, and comparable across time, age and geography. Two layers make it work - a certified score that institutions demand (proctored, protocol-bound, portable) and a practice tier measured by a phone camera every day. A hard schema wall keeps them apart: practice predicts, certification proves.

Both layers stand on one capability: a camera that can score a football drill. Under protocol conditions it certifies a flagship drill; in loose conditions it scores a bedroom drill. Our roadmap calls phone-camera CV \"not a cost-reduction feature but the platform keystone ... both halves of the flywheel depend on it - resource it first.\" This role is that resourcing.

You will author and own the computer-vision R&D programme - the research plan, its milestones, its error budgets - and then run it, from the first labelled rep to models that run offline on a mid-range Android.

Where we are (September 2026)

• A first model is in production. In August 2026 we shipped the first iFreq AI video model to our analyst review platform: temporal action localisation that segments on-ball / off-ball phases and pre-fills annotations for the analysts who collect raw data from drill videos. It covers three drills across seven level variants (BTL L1, DYZ L1-L3, UPL L1-L3).


• A validation layer is being built on top of it - automatic flags for implausible annotations, because analysts get tired and corrupted data must be caught before it reaches a score.


• The platform around the model exists. A video_analysis domain in the backend with a processing queue, stored segment timelines, and automatic metrics kept in their own table so manual and automatic values stay comparable; an on-demand GPU host that wakes when footage is queued and stops itself when the queue runs dry; a pluggable run_inference() seam that is literally waiting for the model code.


• The data is real and growing. Tens of thousands of rep clips, each bound to a player, an exercise, a level and a rep index; per-rep human-entered metrics; and a review timeline that merges analyst and model annotations by authority so unreviewed spans are never silently labelled \"nothing\".


• Scoring is versioned. Every score carries scoring_version and scoring_uid provenance; the coach pentagon is a player-scoped rolling model with fatigue-weighted reps and hybrid z-benchmarks. Your models will feed it, not replace it.

What you will do

Own the research programme

• Write and maintain the CV R&D plan: problem decomposition, milestones, error budgets per drill, build / buy / defer decisions, and the evidence that moves a date.


• Decide what gets modelled and how: temporal action localisation, pose estimation, ball and player tracking, per-rep event detection - chosen against what the protocol actually needs to measure.


• Represent the science externally: reliability and validity publications, the sport-science advisory board, and the federations and academies that will ask \"how do you know?\"

Ship the models (iFreq 2)

• Take the assisted-review model from suggestion to automatic rep detection across all eight exercises of the catalog.


• Build phone-camera timing for the flagship drills and prove it against timing gates under protocol conditions, with an error budget you set and defend.


• Move models to the edge: quantised, on-device inference for the daily session and the free baseline test, with the latency and battery budgets a mid-range Android imposes.

Run measurement science, not just ML

• Design labelling protocols, inter-rater agreement checks and dataset versioning so ground truth is a maintained asset, not a folder.


• Compute and publish test–retest reliability (ICC) per drill, calibration and uncertainty on every automatic value - \"82 ± 3, from 4 certified sessions\" is the product, not a footnote.


• Build the anomaly-defense layer: impossible values, duplicate reps, operator drift, and later cross-angle disagreement - flagged, never silently corrected.

Productionise with the platform team

• Fill the run_inference() seam on the GPU worker and own the model lifecycle behind it: weights, versions, regression tracking across scoring_versions.


• Work with the backend lead on the practice/certified schema wall, with the mobile lead on on-device capture and inference, and with product on what a coach or parent sees when a model is unsure.

Lead the AI track

• Grow the team from one AI engineer to the group the programme needs; set the evaluation bar; review the work.


• Make the work legible to a five-person tech team and a founder-led company: short written updates, decisions with reasons, demos over decks.

You should have

• A research track record in computer vision with depth in video understanding: temporal action localisation or detection, pose estimation, tracking.


• You know the distance between a benchmark number and a deployed one.


• Measurement-science instincts: reliability, agreement, calibration and uncertainty are familiar tools (ICC, Bland Altman, conformal or Bayesian intervals), and you are comfortable publishing them.


• Edge deployment experience: quantisation, ONNX / TFLite / Core ML, and the discipline of a latency budget on a mid-range phone.


• Pragmatic data engineering: labelling workflows, active learning, dataset versioning, and the patience to build the ground truth before the model.


• Deep PyTorch and Python; comfort shipping behind a real API (we run FastAPI, PostgreSQL and AWS) rather than in notebooks.


• Leadership of a small research or applied team, and writing clear enough that an R&D programme can be read and funded on the strength of the document.


• Professional English. French is a strong plus — the team and most of our coaches work in it.

Nice to have

• Sport science or biomechanics: movement screening, athletic testing protocols, youth maturation (PHV, relative-age effects).


• Self-supervised or foundation video models applied to small, domain-specific datasets.


• Experience writing and executing funded R&D programmes (regional or national innovation grants).


• (iFreq 3) Experiment tracking, model registries and drift monitoring in production.

What the first twelve months look like

• By end of 2026: the validation layer is live on the review platform; ground truth is versioned with published inter-rater agreement; auto rep-detection is running as a prototype on the GPU worker.


• Q1 2027: ICC per drill is computed from our own data and published; auto rep-detection measurably reduces analyst time per video on the covered drills.


• Q2-Q3 2027: phone-camera timing for the first flagship drill passes its error budget against timing gates under protocol conditions; the practice-tier scoring proof-of-concept runs in loose conditions.


• By Sep 2027: the R&D plan for on-device inference and the daily session is written, resourced and started.

How we work

Small team, real data, direct line to the founders. Tech is five people plus the CTO; AI is a first-class track with its own ownership and budget, not a service function. Our science is published rather than claimed, our scores carry provenance, and a wrong number shown to one parent is treated as a trust bug, not tech debt.