Building trustworthy data
Creating training data, defining a labeling strategy, and working through class imbalance and distribution shift.
Independent AI project · Work in progress
Case study / 01
Tempo
Building the perception foundation for grounded, near-real-time AI gameplay coaching.
01 / The ambition
Coaching grounded in what actually happened.
Competitive, high skill ceiling games like Overwatch contain an enormous amount of information that could support personalized coaching. The long-term product hypothesis is that a neural network can faithfully extract gameplay observations that are communicated to an AI system that can identify meaningful events, and eventually generate guidance grounded in evidence from the match.
The end product is not plausible-sounding advice. It is coaching that can be traced back to indexed observations of the player's actual game.
Current development milestone
From raw footage to structured observations.
IngestionGameplay videoRaw visual signal
Perception layerComputer vision + MLPerception pipeline
Current outputStructured ManifestGameplay evidence
Active developmentGrounded reasoningIn progress
Next milestoneCoachingNot yet built
02 / Product judgment
Why perception comes first.
Sophisticated downstream reasoning is meaningless if the upstream perception layer is unreliable.
That makes the foundation the priority: turning noisy visual input into structured observations that downstream systems can reason over. The sequencing is deliberate—trustworthy evidence first, richer coaching second.
03 / What exists today
A working ML system—not yet a coaching product.
01
Custom computer-vision and machine-learning pipeline
02
Custom PyTorch model training
03
Recorded and live gameplay inference
04
Structured gameplay-event extraction
05
Purpose-built training data and labeling workflows
06
Evaluation and error-analysis tooling
07
Latency-focused inference experiments
08
Confidence-aware system behavior
04 / The work
What I've been working through.
The difficult part is not a single model. It is building an architecture that is traceable, measurable, and dependable.
Measuring the real system
Reconciling offline evaluation with live-system performance and building tooling that makes errors observable.
Handling uncertainty
Balancing precision and recall, managing confidence, and deciding when probabilistic outputs are reliable enough to use.
Designing for responsiveness
Exploring inference latency and the system constraints that matter when processing feature-rich gameplay in real time.
The central lesson
“Getting a model to work in isolation and designing an architecture that functions without supervision are very different problems.”
The project has become an exploration of what it takes to make probabilistic components reliable enough to support deterministic product behavior: clear evidence, explicit uncertainty, disciplined evaluation, and honest system boundaries.
05 / Product demo
From raw play to a coaching moment.
Six anonymized events reconstructed across fifteen seconds of real gameplay, followed by a simulated example of the coaching experience they could unlock.
Capture process00:00.0
00:00.0 / 00:15.0
Reconstructed event streamLive
Waiting for the first observable signal…
Simulated coaching layerWaiting for full event sequence
Coaching synthesis queued
Watch the full sequence to reveal the review.Product-vision demo built from a real gameplay sequence. Player names are removed.
Product vision · Illustrative outputSee the full coaching experience these signals could unlock.
View coaching report Active development, intentionally private.
Tempo remains under active development. The implementation and repository are intentionally private while I continue exploring the product.