CiteTrace: Knowledge beside the traces

90-second demo

Hiring-manager skim path. Money-shot ask: Why did brake temp spike in S3 on lap 10?

1. Land → Open COTA session
2. Ask: Why did brake temp spike in S3 on lap 10?
3. Click the cite → playhead locks on the evidence band
4. Switch drivers; Compare Soft vs Hard on LEC

Problem → product

Session Q&A without grounding invents physics: a fluent answer that cannot point at a channel window, note, or sector is worse than silence for performance users.

CiteTrace binds every claim to a cite — then locks the playhead and paints an evidence band on the traces. Out-of-window questions fail closed. The wedge is not “AI on telemetry”; it is evidence beside the traces.

What you can do in 90 seconds

1. Open the COTA session — LEC Soft flyer by default
2. Use Lap for the whole lap, or S1–S3 for sector focus; scrub the playhead
3. Ask a grounded question → claim + cites → click a cite → lock
4. Switch drivers; Compare vs Soft flyer / Hard baseline

System shape

One session pack, three stores, one retrieval path. Cites are first-class objects — not chat footnotes.

19 notes · 13 events · 27 cites · 15 channel docs

Engineering judgment

• Deterministic QaMap — the money-shot demo never flakes; honesty over LLM theater for a portfolio walkthrough.
• OpenF1 subset + labeled overlays — real broadcast car_data, pits, weather series, and team radio for the field; synthetic rear/front brake temp, brake pressure, tyre surface, steering, long-g, fuel flow, oil, and coolant where public feeds have no ATLAS-class plant channels — provenance is explicit in the chrome.
• Distance-domain stage — whole-lap and sector focus, playhead scrub, compare overlays; matches how engineers argue about the lap.
• Field + depth — best flyer for all 22 cars; LEC Hard / Soft / long-run trio for Soft-vs-Hard narrative depth.

Craft signals

• React / TypeScript UX — dense pit-wall chrome, SVG waveforms, keyboard-focusable controls.
• Multi-store literacy — samples, session metadata, and notes as one grounded corpus.
• Telemetry vocabulary — session, lap, channel, compare, notes, events — shaped like public ATLAS docs without cloning vendor UI.
• Provenance labeling — OpenF1 vs synthetic overlay called out in-product; fail-closed when the window cannot support the claim.

Stack

Data honesty

OPENF1 SUBSET · 2026 US GP FP2 (COTA) · synthetic ATLAS overlays · not SECU

Pack cota-2026-fp2.json —22 drivers (best flying lap) plus LEC Hard/Soft/long-run. Real OpenF1 car_data, location path, weather, stints, race control. Overlays: brake_temp, brake_temp_fl, brake_pressure, tyre_surf_fl, steering, g_long, fuel_flow, oil_pressure, coolant_temp. Regen with npm run fetch:cota. Not ATLAS software.

What we do not fake

• Employment, offers, or partnerships with any team or vendor
• Live SECU, paddock RF, factory VPN, or real team channel maps
• Vendor Viewer UI clones, logos, or Automation / SQL Race credentials
• Kafka/Docker as a runtime dependency for this portfolio slice
• Full redistributed FastF1 caches, OpenF1 session dumps, or iRacing IBT samples

• OpenF1 — public API (https://openf1.org/)
• Circuit of the Americas — 5.513 km public circuit facts
• ATLAS public docs — apps/parameters vocabulary for overlays only
• Shaped like: OpenF1 car_data / laps / location / weather / stints / race_control
• Shaped like: ATLAS apps/parameters (synthetic thermal overlay)
• Shaped like: OpenF1 pit / weather / team_radio