- Project number
- Project / 001
- Year
- 2026
- Category
- Product
- Status
- Active · Live
Sepang Vision Lab
A 3D, broadcast-style replay of the 2026 Sepang weekend. Every practice session, qualifying and the race, rebuilt from public OpenF1 data: real names, positions, telemetry, tyres, race control and weather, with the cars driving an aligned 3D model of the circuit.
- Status
- Live
- Role
- Solo: product, data pipeline, 3D, engineering
- Timeline
- 2026 · 29 milestones
- Data
- OpenF1 · 2026 Sepang weekend (2 to 4 October) · 5 sessions, 22 cars
- Stack
- React 19 · TypeScript · Three.js (React Three Fiber) · Vite · Python · Claude API
- Live
- sepangvisionlab.madebyrazeen.com
01 /Overview
Watch the race again, from any seat.
Sepang Vision Lab replays the 2026 Sepang weekend in 3D, in the browser. Pick a session, follow any of the 22 cars and switch between five cameras while a timing tower, track map, weather and race-control messages run alongside, the way a television broadcast would show them.
Everything on screen comes from recorded data. Where the data cannot say something, such as track width, brake temperature or the scenery around the circuit, the app either leaves it out or labels it as illustrative.
02 /The problem
Public data is sparse and noisy.
OpenF1 publishes car positions at about four samples a second, with jittery timestamps, in a coordinate system that matches no map. Tyre stints sometimes contradict the pit stops. Sessions have gaps.
Played back as they arrive, the cars stutter, cut corners, overlap and float above the track. The work was turning that into motion that looks physical without inventing anything the data does not contain.
03 /Gameplay
What it looks like.

Race, lap 11, under the safety car: the TV camera on Charles Leclerc, the 22-car timing tower, the driver card with live speed, gear, throttle and g-force, and the replay bar with incident markers.
Real product · 2026 Sepang race replay, recorded OpenF1 data
On a phone: one scrolling column with the 3D view, replay controls and the followed driver's telemetry.
Real product · 2026 Sepang race replay, recorded OpenF1 data
Chase camera with the driving HUD: rev arc, gear and g-meter, all from recorded channels.
Real product · 2026 Sepang race replay, recorded OpenF1 data
Inspect camera around a team-style livery. Colours only, no logos.
Real product · 2026 Sepang race replay, recorded OpenF1 data
The pit lane, derived from where cars actually drove during the race's 73 pit stops.
Real product · 2026 Sepang race replay, recorded OpenF1 data04 /System architecture
Fetch once, replay in the browser.
05 / OUTPUTReact Three Fiber scene reading the replay clock every frame.
05 /Data / intelligence
Making noisy data move like a car.
Alignment
OpenF1 positions are fitted to the circuit model with a derived similarity transform: 3.09 m RMS and 5.69 m at the 95th percentile from the centre line, a residual that includes the real racing line.
Smooth motion
Each car's distance along the track and its offset across it are refitted with a Gaussian-weighted local fit against time, so cars follow their real line and steer from its curvature.
Physics check
Sampling every car at 50 Hz across the race, 99% of frames stay within 3.7 g braking or acceleration and 5.9 g cornering, the range of a real car.
Elevation
About 22 m of climbs and drops, derived from OpenF1 heights. Cars, cameras, scenery and terrain all follow it.
Cars keep their space
Aligned positions carry a few metres of error, so overlapping cars are eased apart sideways. Overlapping frames in the race fell from 3,955 to 211, and only pairs that race control reports as colliding may touch.
Honest gaps
When OpenF1's tyre labels disagree, the tower shows ? instead of a guess. A data gap holds the last sample and marks the car stale after two seconds.
Gap between samples · 6,830 samples
Median gap 240 ms, about four samples a second, and 54 gaps longer than half a second. Played as received, the car stutters.
Car 1 · Race, 4 Oct 2026 · OpenF1 · smoothing here is a 0.5 s time window for illustration; the app fits motion along the track
06 /Build
Broadcast features on top.
01Five cameras
TV, Chase, Onboard, Heli and Inspect, with a speed-sensitive lens and a driving HUD in Chase and Onboard.
02Compare and share
Compare two drivers on the same lap with an optional ghost car, and copy a link that reopens the exact session, moment, driver and camera.
03Weather
Rain falls while OpenF1 reports rain, then the track dries over about 90 minutes, a drying time fitted to the race's intermediate stint.
04Race engineer
Ask about gaps, tyres, pace or flags and hear the answer as a radio call. Presets are answered from the data with no key. With the viewer's own Anthropic API key, Claude answers anything from the same data brief and is told to use only its numbers.
05Hand gestures
Optional webcam control through MediaPipe, running locally: swipe to skip, pinch to follow the next driver. No frames leave the browser.
06Fast first load
three.js and MediaPipe are kept out of the first download, which fell from 988 KB to under 300 KB.
07 /Outcome
Live, and checked against the source.
- Sessions
- 5
- FP1, FP2, FP3, Qualifying, Race
- Cars
- 22
- every car in every session
- Alignment
- 3.09 m
- RMS to the centre line
- Tests
- 125
- all passing
All five sessions were compared with the live OpenF1 API for laps, stints, pits, race control and classification, and the position and telemetry streams have no gap over 60 seconds.
Earlier versions also carried lap-time models and strategy tools. Those were removed so the project could focus on doing one recorded weekend well.
08 /What I learned
Data first, then the polish.
01Fix the data before the visuals
Most of what looked like rendering bugs, such as stutter, cars overlapping or floating, came from the data. Fixing it at the source fixed every view at once.
02Label every number
DERIVED, SOURCED, OFFICIAL or illustrative: saying where each value comes from is what lets a replay be trusted.
03Cut to focus
Removing the strategy and ML tools made the product clearer, not smaller.
04Give AI a narrow job
The race engineer only answers from a brief of recorded data, so it explains the race without inventing it.
Read the journal
Making Noisy F1 Data Move Like a CarIndependent motorsport technology and research project. Motorsport names and trademarks belong to their respective owners. No official affiliation or endorsement is implied. Data via OpenF1 (unofficial). Driver and team names are shown for identification only, and no team, series or sponsor logos are used.
09 /Next project
BalangA Malaysian probability party game: read the jar of 25 kuih tokens, bet on secret prediction cards, and play friends in peer-to-peer rooms with no backend.
