Technology Overview

KiSTI combines motorsport-grade sensors, a production ECU, and edge AI into a unified telemetry platform. Here's how data flows from sensor to insight.

System Architecture

Sensors→CAN→Link→Jetson AGX Thor→
Driver View
ALDCZeus→Pit Engineer

Data Pipeline

1. Sensor Layer

13 high-frequency sensors (brake temps, tire temps, EGT, boost, oil temp/pressure, wideband O₂) sample at 4-10Hz via analog/CAN interfaces.

Brake FL/FR/RL/RR thermocouples, tire FL/FR/RL/RR infrared sensors, K-type EGT probe, Cobb 4 Bar MAP sensor, GM IAT sensor, 150 PSI oil pressure sensor, flex fuel sensor, Bosch LSU 4.9 wideband.

2. Vision Layer

4 front-mounted cameras feed directly to the Jetson AGX Thor via USB 3.0 and CSI — thermal, depth, visual, and weather sensing.

Teledyne FLIR thermal IR, 3D LiDAR point cloud, high-speed RGB camera, Yoctopuce Yocto-Spruce weather station.

3. ECU Aggregation

Link G5 Neo 4 ECU receives all sensor data via CAN bus, applies calibration tables, and streams merged telemetry over USB.

500Kbps CAN, 100+ channels available, configurable output rates, real-time fuel/ignition corrections.

4. Edge Inference

NVIDIA Jetson AGX Thor processes telemetry and vision data at the edge — anomaly detection, pattern matching, and predictive diagnostics in <50ms.

1,000+ TOPS AI performance (128GB LPDDR5X), TensorRT optimized models, 16 camera inputs, local data buffering when offline.

5. Cloud Sync

Zeus ingests telemetry via WiFi/cellular with store-and-forward. AI findings surface in real-time dashboards.

pgvector semantic search, 3.5M+ memories, automatic embedding via Voyage AI, Slack/email alerts.

Voice AI Pipeline

Fully offline conversational AI — no cloud dependency, no cellular required. The entire voice stack runs on-device on the Jetson AGX Thor.

Speech-to-Text

whisper.cpp with CUDA acceleration

~130ms for 3-second utterances. Custom wake word detection with barge-in echo cancellation.

Text-to-Speech

Piper TTS — sub-200ms conversational responses

Three voice modes: Informal (personality), Standard (clinical data), Safety-Critical (emergency only).

Edge Memory

DuckDB + ONNX embeddings on-device

Persistent knowledge across sessions. Voice-activated 'remember' commands. Cloud sync when WiFi available.

Anomaly Detection

Real-time pattern matching on telemetry streams

Threshold-based alerts sourced from build baselines. Predictive diagnostics with spoken warnings.

The Vehicle

2014 Subaru WRX STI Hatch. IAG 750 Closed Deck short block (rated 750 BHP), BCP X400 turbo, ID1300 injectors, COBB front-mount intercooler. Full standalone flex fuel tune by Boost Barn Motorsports. Track-tested at Mission Raceway, BC.

E85 Corn EthanolIAG 750 BlockBCP X400 Turbo360-390 WHPLink G5 Neo 4 ECUFortune Auto CoiloversCompetition Clutch Stage 2

Runs exclusively on E85 — corn-derived ethanol. One of the few edge AI automotive platforms powered by renewable fuel.

17

Sensor Channels

4-10 Hz

Sample Rate

<50ms

Edge Latency

500 Kbps

CAN Speed

1,000+ TOPS

AI Performance

3.5M+

Cloud Memories