NADIR Core

Open-source ADAS drift scoring, plus a free local dashcam agent that estimates camera mount / geometry health and scores it with the same Pulse engine.

Version 0.2.0 (beta). Apache-2.0. Shadow mode only — no ECU writes.

Repo: https://github.com/thedhruvhegde/nadir-core · Product: https://nadirai.net

Pulse scoring works today on numpy telemetry. The dashcam lane works end-to-end on synthetic frames and on folder / VIOFO / BlackVue adapters (live hardware depends on your firmware and Wi‑Fi). Novel residuals (IGMR, LI-CUSUM, soft coupling) are implemented with writeups under docs/proofs/. See docs/STATUS.md and docs/TESTING.md.

python -m venv .venv && source .venv/bin/activate
pip install -e ".[dashcam,dev]"
pytest -q
nadir-dashcam analyze --source synthetic --yaw-drift 1.2

git clone https://github.com/thedhruvhegde/nadir-core.git
cd nadir-core
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dashcam,dev]"   # scoring + dashcam
# or: pip install -e ".[dev]"     # scoring only

Needs Python 3.9+.

nadir-core-score --yaw-deg 0.35

from nadir_core import SensorReadings, score_pulse
import math

def rot(yaw):
    r = math.radians(yaw); c,s = math.cos(r), math.sin(r)
    return [c,-s,0, s,c,0, 0,0,1]

print(score_pulse("demo", SensorReadings(camera_rotation_matrix=rot(0.4))).tier)

If you have a VIOFO A229-class camera (common on Amazon US), a synced SD folder, BlackVue local Wi‑Fi, or RTSP:

nadir-dashcam connect --brand viofo
nadir-dashcam analyze --source synthetic --yaw-drift 1.5 --store out/demo.jsonl
nadir-dashcam analyze --source folder --path ./recordings
nadir-dashcam analyze --source viofo --host 192.168.1.50
nadir-dashcam analyze --source blackvue --host 10.99.77.1
nadir-dashcam report --store out/demo.jsonl

Hybrid behavior: analyze live when a stream is up; otherwise score clips as they land after parking. Video stays on disk unless you pass --upload (score JSON only).

Full schematic: docs/pipeline/SCHEMATIC.md.

Short version: frames → horizon / vanishing / flow → EKF → IGMR + LI-CUSUM + soft coupling → score_pulse → JSONL / CLI.

We ship three NADIR-named formulations for the dashcam lane:

1. IGMR — information-geometric mount residual (Fisher–Rao + SO(3) guard)
2. LI-CUSUM — left-invariant CUSUM on SO(3) innovations
3. SCE — SPD soft-coupling energy into a Pulse-like scalar

Propositions and proofs: docs/proofs/. Implementation: nadir_core.dashcam.math_novel.

Classical pieces (RANSAC VP, pyramidal LK, Huber IRLS, Welch PSD) sit underneath. None of this is an ASIL proof.

Dashcam pixels do not invent radar/lidar OEM residuals. This lane is vision mount / geometry health bridged into Pulse. Not a calibration certificate. Not ASIL.

In: Pulse/Echo scoring, dashcam agent, synthetic generators, schemas, proofs, CI.
Out: Console, auth, digests, signed evidence product, customer data, transit/OEM packs.

CONTRIBUTING.md. Useful PRs: brand adapters, scrubbed fixtures, numerical tests, firmware-specific VIOFO notes.

Apache-2.0 — LICENSE.

founders@nadirai.net · https://nadirai.net/technology/validation