Open Source · IAI09 Capstone · by Datronex

Deeper diagnostics,
plain language.

Your vehicle generates rich sensor data on every drive. MisfireAI reads it, scores it, and tells you what it means — before a warning light ever comes on.

View on GitHub See How It Works

Standard code readers tell you what broke.
MisfireAI tells you what's breaking.

OBD2 was designed for emissions monitoring, not predictive diagnostics. Most tools surface fault codes after damage is done — leaving owners dependent on mechanics with no way to verify the diagnosis.

🔌

Rich data, no interpretation

Every 1996+ vehicle continuously broadcasts dozens of live sensor readings. Most owners never see them, and tools that do show them offer no context.

⚠️

Codes appear after the fact

A fault code means the ECM already confirmed a problem. The sensor trends that led there — fuel trims creeping over weeks, catalyst efficiency declining — go unnoticed.

🔧

No independent verification

When a mechanic quotes a repair, there's no accessible tool for the average owner to verify the diagnosis against their own vehicle's data.

Four stages. One pipeline.

MisfireAI follows a structured signal pipeline — from raw OBD2 data to a plain-language diagnostic assessment with human approval at every high-stakes step.

① Catch

Ingest

Parse OBD2 log files from any source — Car Scanner, MHD, Techstream, or raw CSV. Normalize to a common PID schema regardless of format.

② Enrich

Contextualize

Decode the VIN via NHTSA to identify the exact vehicle. Look up open recalls and complaints. Add vehicle-specific context before analysis.

③ Separate

Score

Score each vehicle system — fueling, cooling, ignition, catalyst — on a 0–1 scale. Separate sustained anomalies from transient noise using four severity tiers.

④ Compound

Advise

GPT-4o generates a plain-language diagnostic assessment. High-stakes repair briefs require explicit owner approval before any action is taken.

👤

Human approval on every repair brief

When the pipeline flags a high-severity finding, it doesn't act automatically. The vehicle owner receives an email with the full assessment, system health scores, and any active fault codes. They approve or reject. The decision — and its timestamp — is logged regardless of outcome. No repair recommendation is forwarded or stored without an explicit owner action on record.

Built for real vehicles, real data.

Works on any 1996+ US vehicle with a standard OBD2 port.

🩺

Per-system health scoring

Fueling, cooling, ignition, and catalyst each scored 0–1. See exactly which system is degrading and how far it's drifted from normal.

🔍

Signal vs. noise separation

Four-tier anomaly classification separates a single transient spike from a sustained pattern developing across a drive cycle.

🪪

VIN-aware analysis

Vehicle identity decoded via NHTSA. Open recalls and complaints checked automatically. Analysis adapts to the specific make, model, and engine.

📊

Full observability

Every pipeline stage traced in Phoenix/Arize. Every tool call, every score, every approval decision logged with timestamps and attributes.

🔌

Multi-source ingestion

Accepts log files from Car Scanner, MHD, Toyota Techstream, carOBD, and generic CSV formats. Auto-detects source format and normalizes column names.

📦

Portable open-source stack

Clone the repo, add your API keys, run the pipeline. Sample data included. No proprietary infrastructure required to get started.

Real data from real vehicles.

Pipeline validated against log files from multiple makes, models, and engine types.

2009 BMW 335i — N54 twin-turbo
2007 Toyota Tundra — 5.7L 3UR-FE
2015 Honda Fit — 1.5L L15B7
Toyota Etios 2014 — 1.5L 2NR-FE
KIA Soul — 1.6L G4FD
Multi-make Brazil fleet — cephasax dataset

Request a demo or ask a question.

Interested in MisfireAI for your vehicle, fleet, or shop? Reach out directly.

misfire@datronex.net