Skip to main content
APEX

IBM SkillsBuild · May Challenge 2026

The race engineer for the drivers who don’t have one.

A pro race engineer can cost several hundred poundsa day. Every F1 driver has one. Most adaptive racers, veteran-team drivers, and grassroots competitors do not. APEX is built on IBM Granite, the same platform IBM ships to Scuderia Ferrari’s ~400 million fans, and points it at the drivers who need a race engineer the most.

Built for race engineers, drivers, adaptive-racing coaches, and grassroots programs.

See how it worksUpload telemetryCoach my codeTwin-driver compare

IBM Granite + watsonx · Apache 2.0

Editorial paddock motorsport illustration: hand-controls-adapted GT4 cockpit interior at golden hour. Warm cream paper textured backdrop with subtle racing-green track-line silhouette. Clay-red apex marker. Amber heat-glow on the dashboard telemetry. Generated via Nano Banana Pro per docs/nano-banana-runbook.md wave-51c ship.

Who APEX races for

Adaptive racers, veteran-team drivers, grassroots competitors. The people who need a race engineer the most.

APEX is built first for drivers competing under FIA Appendix L hand-control certification: differential thumb-trigger paddles, modified pedal assemblies, and brake-throttle simultaneity envelopes that pro-grade coaching software silently treats as data corruption. Then the same coaching surface generalizes to any racer who cannot afford a pro engineer at £400 to £500 a day.

  • Control input

    Hand-controls preserve continuous-action capability.

    Most pro tools assume brake + throttle never overlap. For adaptive drivers using thumb-triggers, simultaneous activation IS the technique. APEX’s COA-gate inverts the assumption.

  • Regulation

    FIA Appendix L grounds every coaching claim.

    Granite Guardian audits each recommendation against Appendix L provisions. No invented Article numbers. Citations point at the published revision, never at a hallucinated reference.

  • Community

    Co-developed with operators who actually field these cars.

    MME Motorsport (per consent 2026-05-22) reviews the COA-parameterized gate behavior against live setup data. Additional operator partners anonymized per per-surface consent discipline.

  • Open source

    Apache 2.0 from Day 1. Public repo, public methodology.

    The race engineering canon should not be locked behind a paid SaaS login. APEX ships every layer (forecaster + projector + Guardian) as inspectable open code under a permissive license.

“The race engineer is the unsung half of a Sunday result. Every paddock has one in the truck. Almost nobody outside that paddock does.”

What the driver brings

Three artifacts, one report.

Telemetry CSV from the data logger. FIA Certificate of Adaptations PDF when the driver has one on file under FIA Appendix L provisions. A sentence or two from the driver’s own debrief. APEX returns a corner-by-corner coaching report inside five minutes.

See the canonical demo at /judges. Try with your own files at /analyze.

Lost the rears mid the slow hairpin again. Trail-braking on the lever doesn’t come back the way it did last month. The middle sector was plus zero point three against my PB.

Illustrative debriefcorner-agnostic example+0.34s mid-sector delta

The COA gate, visualized

Brake-throttle traces crossing at the apex. When the COA permits it, the projection recognizes it.

Editorial paddock infographic: clay-red brake telemetry trace + racing-green throttle telemetry trace overlaid through a corner, crossing at the apex point. Amber heat-glow halo around the apex crossing. Warm cream paper background with subtle paddock dust grain. Plex Mono callout coa_overlap_flag = 1. Generated via Nano Banana Pro per docs/nano-banana-runbook.md wave-51c ship.
Editorial illustration. coa_overlap_flag is the tensor-level signal the V1 NumPy validator reads from the parsed FIA Certificate of Adaptations.

PhysicsTTM, three layers

We didn’t train a custom PINN.
We constrained an off-the-shelf TSFM.

TTM was pretrained on weather and retail. Without constraints it can forecast 4G lateral with zero steering, or speed climbing with the throttle at zero. APEX inserts a differentiable physics layer between the forecaster and the report. To our knowledge no prior published work does this for vehicle dynamics without retraining the model from scratch.

  1. 01layer

    Frozen forecaster

    Granite TimeSeries TTM r2.1

    A pretrained, sub-1M parameter time-series foundation model. Outperforms several larger TSFMs in the NeurIPS 2024 benchmarks on common forecasting tasks. We do not retrain. We aggregate raw 50 Hz telemetry to 1 Hz mini-sector tensors that sit inside the model’s published support envelope.

    frozen · zero-shot · channel-independent

  2. 02layer

    Differentiable physics projection

    CvxpyLayers QP

    Every forecast step passes through a QP that enforces the friction ellipse, the bicycle model, a forward-Euler kinematic step, a jerk bound, and a circuit-conditional friction lookup. Plus the COA simultaneity flag. When the driver’s Certificate of Adaptations permits simultaneous brake and throttle, we permit it. Otherwise enforce.

    differentiable · COA-aware · per-step μ_v (Tier 5 thermal + Tier 7 Pacejka) per D-015

  3. 03layer

    Safety-classifier text audit

    Granite Guardian 4.1 8B (BYOC)

    The projection layer emits a structured English text log of every constraint correction. Guardian audits the log against custom Bring-Your-Own-Classifier rules. A unit-test suite covering every kinematic violation type backs the serializer (Convergence 14, our load-bearing safety contract). Reasoning trace surfaces in the UI.

    text audit · serializer-tested · think-mode visible

What makes it different

Five claims. Each independently verifiable.

  1. 01

    First pretrained time-series foundation model on motorsport telemetry.

    vs Deep Dynamics, which trains a bespoke PINN from scratch; vs Chronos-on-car-following, which uses a different forecaster on different inputs. We freeze TTM and wrap it.

  2. 02

    First AI to ingest the FIA Certificate of Adaptations as a binding safety envelope.

    FIA Appendix L is the regulation that governs Certificate of Adaptations structure for adaptive-equipment homologation. APEX parses the COA at onboarding, derives the c_overlap flag from the approved hand-control hardware specifications inside it, and feeds the flag to the model at the tensor level. Not regulatory background; tensor-level input.

  3. 03

    First integrated workflow for adaptive driver hand-control channels.

    COA-parameterized brake-throttle simultaneity. Competing tools hard-code a brake-throttle mutual-exclusion and systematically misdiagnose adaptive drivers, and any left-foot-braker.

  4. 04

    Granite Guardian with BYOC custom rules, backed by a serializer unit-test suite.

    Convergence 14: every kinematic violation type has a fixture text log and a verified Guardian verdict. Safety-critical code with safety-critical test coverage.

  5. 05

    Fourteen IBM Granite tools across a per-tool honesty ladder.

    Three WIRED at HEAD, nine at INTEGRATION with canonical type contracts + backend swap-points, two build-time accelerators (Docling library + Mellea v0.5.0 IVR-loop slot; Mellea is build-time architectural inspiration only, not a runtime dependency). Per the IBM × Scuderia Ferrari case-study precedent.

IBM stack

Fourteen tools. Per-tool wire-up status, honest.

Three tools wired end-to-end at HEAD (Granite Instruct 4.1 8B coaching narration + Granite 4.0 Nano 350M WebGPU edge model + Granite Embedding R2 live dense RAG rerank via HF Inference Providers). Nine tools at frontend-integration phase with canonical type contracts + backend swap-points documented (per Stream M.3 spec handoff + wave-46 D-058 expansion adding Granite Instruct 4.1 3B chat-routing + Granite Speech 4.1 2B-Plus Watson STT proxy preview; LangGraph + Granite MCP Gateway + ContextForge as the orchestration runtime per D-017 G7 + D-054; render path stays identical across mock and real; Langflow retained as the export-graph artifact). Two build-time accelerators (Docling library + Mellea v0.5.0 IVR-loop slot; Mellea is build-time architectural inspiration only, not a runtime dependency). No runtime tool listed without a runtime role in the pipeline.

  • Granite-DoclingIntegration
    258MFIA COA PDF to structured JSON parser
  • Docling libraryAccelerator
    latestOpen-source IBM Docling conversion + table-extraction
  • Granite VisionIntegration
    4.1 4BSRO + Britcar timing-sheet PDF to CSV
  • Granite TimeSeries TTMIntegration
    r2.1 + fine-tuneTrack 1 of D-010 three-track ensemble (frozen-backbone + channel-mix decoder fine-tune per D-052)
  • Granite FlowStateIntegration
    r1.1 18.5MTrack 2 of D-010 (sampling-rate-invariant continuous-time SSM at 50 Hz)
  • IBM TSPulseIntegration
    1MPolyphase time-frequency anomaly detector (D-016 Layer 2)
  • Granite Embedding R2Wired
    149M + 47MLive dense RAG rerank via HF Inference Providers (granite-embedding-30m-english feature-extraction); lexical fallback (D-016)
  • Granite InstructWired
    4.1 8BRace-engineer narrator: live /analyze coaching report + chat via OpenRouter (wave-64; verified phase=real in prod)
  • Granite GuardianIntegration
    4.1 8BBYOC custom-rule audit + D-024 physics-confidence downgrade
  • Granite 4.0 NanoWired
    350MIn-browser WebGPU edge model via Transformers.js (D-019 + D-021)
  • LangGraph + MCP + ContextForgeIntegration
    runtime (Langflow export-graph)Orchestration runtime per D-017 G7 + D-026 (Langflow demoted to export-graph facade per wave-30)
  • Granite Instruct 3BIntegration
    4.1 3BFast-path AICopilotChat routing for simple queries (D-058 wave-46 add)
  • Granite SpeechIntegration
    4.1 2B-PlusSpeaker-attributed ASR + word-level timestamps replacing Web Speech (D-058 wave-46 add)
  • MelleaAccelerator
    >=0.5.0Instruct-Validate-Repair tri-agent critic loop scaffold on narrator (IBM Research open source library, Apache 2.0); flips to WIRED when Vinh M3-V5.2 ships Mellea-orchestrated narrator path