Tesla Intelligence Core v2.0.0-preview.1

GHOST V6.6 local AI, verified memory & voice alerts — development preview

PRE-RELEASE — NOT A COMPLETE INSTALLABLE V2 BUILD.
The current developer Raspberry Pi runs the larger Qwen3.5 4B explanation engine, but this GitHub revision still uses the public v1.0 installer. Installing or upgrading from this preview will not recreate the entire V6.6 intelligence stack. Do not run it over a working developer installation.

What's new on the current development GHOST

Qwen3.5 4B via local llama.cpp: optional asynchronous, cached natural-language analysis explaining verified Tesla charging, pack temperature and separate readiness scores.
Evidence-validated AI: numeric/relational guards reject unsupported claims, false module-spread comparisons, unwarranted sustained trends, invented numbers and unjustified fault-free claims; deterministic verified text is the fallback.
Voice & smart alerts: browser spoken verified readings, optionally spoken AI analysis, on-the-hour updates and independent rule-based thermal/charging/status alerts.
V5 historical Tessie insights + V6.2 verified hourly memory: charging-power and module-temperature historical context, direct battery-heater/HVAC separation and local timestamped observations. Memory retrieval is not language-model weight retraining.
Readiness dashboard expansion: distinct RIGHT NOW and NEXT DEPARTURE scores, live battery/heater/cabin state, pack outlook, Open-Meteo outdoor forecast, and vehicle-reported versus derived pack power.
Shadow reliability improvements: Qwen model handoff waits for active briefings to finish; current 4B has a watchdog health check and a retained smaller model rollback configuration. Full autonomous crash/restart/rollback behavior is still under test.
Ambient thermal model research: 90-day descriptive analysis and chronological holdout of qualified parked windows. The experimental ambient predictor performed worse than the simple baseline and was not promoted.

Genuine latest screenshot

See the actual current development dashboard

Source and installation status

This preview is meant to document and show the significant development update, not to replace the user's running Pi installation.
The existing GitHub installer still targets v1.0.0. The latest 4B model/voice/source modules and services are being reconciled in PR #4.
A private source review bundle was obtained from the live Pi; it is not published unreviewed because it contains environment-specific paths and review-only IP redactions.
No GGUF weights, personal vehicle readings, private routes, API tokens or environment files are included.
The full V2 upgrade will follow after source migration, portable service templates, model dependency installation, security review, clean Raspberry Pi installation, and rollback testing.

Further details

Complete changelog · Feature matrix · V2 architecture · Full upgrade notes · Blocking release gates

Assets 2

Tesla Intelligence Core v1.0.0 — Winter & Whole-Vehicle Intelligence Latest

Latest

github-actions released this 16 Oct 02:15

v1.0.0

3919ae9

Tesla Intelligence Core v1.0.0 — Major release

Public major release with a hardware-test limitation. Development-system usage, source checks and automated regression tests have been completed, but a completely new Raspberry Pi installation and on-device upgrade/reboot regression were not performed before publication. This limitation is disclosed so new users can decide whether to try the installer; do not treat it as independently certified production software.

What this project does

Tesla Intelligence Core started as a way to understand overnight winter cold-soak, battery heating, Level 1 AC charging, and whether the car would have enough battery for the next drive. It has expanded into a local-first, read-only intelligence platform that learns how a specific Tesla behaves over time. All completed drives can contribute useful data, not just regular commutes; unfamiliar destinations and weekend errands are not forced into a home/work assumption.

v1 release scope

Thermal intelligence: future pack-temperature outlook, outside-weather forecasts (3/6/12 hours), learned parked thermal retention, warm-up and battery heater context, module temperatures and morning-readiness projection.
Charging intelligence: learns slow AC charging and separates fast-charging events from low-power charging evidence. Tracks limits, thermal auxiliary demand and actual input where supported.
Vehicle intelligence: all-trip energy extraction and event history; multiple shadow stages for trip sequence, site context, GPS-confirmed outcomes, feedback, time-only coverage, and trip relationships.
Forecast accountability: records predictions before real events and checks outcomes, with distinct score units and explicit coverage limitations. Forecast Quality v2 and Trip Quality v3.2.1 run hourly, independently of model training.
Vehicle sensors: read-only TPMS values, vehicle lock, Sentry Mode, climate fan, charge port and other available state fields. TPMS values are converted from Tesla bar units to PSI; staleness is shown.
Machine learning: classical baselines, neural challenger, governor and held-out checks. Shadow evaluators do not automatically alter the serving model.
Dashboards: existing Winter Readiness Home unchanged in layout, separate Forecast Quality, Trip Quality and Vehicle Sensors pages; dark-theme real-time presentation.

v3.2.1 charging-aware evaluation

A significant drive after a frozen forecast may invalidate a comparison. However, a charging session already running when that forecast was captured is part of known baseline conditions, not automatically a disqualifying interruption. A new charge initiated after the capture remains withheld unless reliable evidence establishes the forecast explicitly planned it. This does not prove the charging amount was predicted correctly: departure SOC error remains measurable separately.

Every trip contributes to the observational history even when no suitable prior forecast exists. A skipped or withheld forecast is not counted as a model success or model failure. Time-of-day labels are never proof of the destination.

One-command setup design

Supported target platform: Raspberry Pi 5 with Debian / Raspberry Pi OS 64-bit, Python 3.11+.

bash <(curl -fsSL https://raw.githubusercontent.com/prokyle123/tesla-intelligence-core/main/install.sh)

This command targets the public v1.0.0 release after the release branch is merged. Use a current backup before upgrading an existing installation.

A first-time user answers a few guided prompts:

Linux service user and local timezone (critical for departure learning).
One telemetry source: Tessie API token and optional VIN or existing TeslaMate MQTT host, car ID, username, password and TLS.
Optional departure seed (fades as historical evidence grows), Tessie historical backfill duration and dashboard port.
Optional EcoFlow River, Starlink status or Tailscale access. No optional external hardware is required for the core features.

The installer should then create private data/config directories, install Python dependencies and portable systemd services, initialize database schema, start telemetry collection, schedule model training, forecast-quality and trip-quality verification, and shadow trip learning. Credentials stay on the user's own Pi, not GitHub.

An existing installation should preserve its configuration, credentials, telemetry history, model versions, scores and private trip context. Upgrading an existing installation: Back up your configuration, database, learned model artifacts and local shadow sidecars before updating. The installer is designed to preserve those, but the upgrade path has not yet been validated on another live system.

Portable background services

Unit/timer	Purpose	Intended schedule
`ghost-tesla-ai-collector.service`	Live telemetry collection	Continuous
`ghost-tesla-ai-web.service`	Dashboard and read-only APIs	Continuous
`ghost-tesla-ai-intelligence.timer`	Event intelligence	Existing system schedule
`ghost-tesla-ai-train.timer`	Baseline training	Existing system schedule
`ghost-tesla-ai-neural.timer`	Neural challenger/governor	Existing system schedule
`ghost-forecast-quality-v2.timer`	Observed context-sensitive forecast error	Every hour
`ghost-trip-quality-v32.timer`	Whole-journey forecast scoring	Every hour
`ghost-trip-shadow.timer`	Six existing shadow learners in dependency order	Every 30 minutes

Private registry, GPS sidecar, quality reports, and model evidence never belong in release assets. Site IDs do not imply user-identifiable routes.

Data accuracy / transparency

Winter Readiness Score is an operational heuristic, not a calibrated model-accuracy percentage.
Time-only departure labels, GPS-verified route outcomes, and complete candidate journeys are distinct concepts.
Minute-by-minute rolling windows are correlated. Use independent episodes and completed journeys for model comparison; show sample counts.
Weather forecasts may use Open-Meteo; no vehicle controls or safety-critical commands are sent.
Battery energy and Wh/mile should be measured against supported telemetry evidence; do not assume all charging-wall input reaches the pack.
A missing field, stale observation, or delayed collector is shown as missing/stale rather than silently substituted with a confident prediction.

Release validation matrix

Release support statement: v1.0.0 is a public, open-source major source release with automated checks; independent Raspberry Pi deployment verification remains outstanding. Contributions and compatibility reports are welcome.

Assets 2

Tesla Intelligence Core v0.8.27.6

prokyle123 released this 28 Sep 20:15

v0.8.27.6

0196794

First public release of Tesla Intelligence Core. Local-first Tesla telemetry, winter-readiness, thermal, charging, trip-pattern and neural prediction system for Raspberry Pi.