A Pi 5 sensor body
Vision, audio, display, playback, face crops, pose, expression, and scene summaries stream from a Raspberry Pi 5 with Hailo acceleration.
Private intelligence. Local hardware. Earned autonomy.
A personal AI that grows with you.
JARVIS Oracle Edition is a locally-hostable, fully-governed cognitive substrate with self-directed skill acquisition, self-modifying code under quorum, an epistemic immune system, internal neural dialog, and restart-honest continuity.
Not a cloud wrapper. Not a disposable chatbot. A private Oracle that senses, remembers, studies, verifies, and matures on your own machines.
integrative
memory reflection · policy shadow · epistemic integrity
private local intelligence, awake on your own hardware
architecturally complete · runtime maturing
JARVIS ORACLE
100%local-first core runtime
13epistemic integrity layers
11distillation specialists
L2safe-apply autonomy lane
GoldOracle benchmark seal
JARVIS in motion
Drop in screenshots. Let the interface detonate into view.
01 / 05
Operator view
The Oracle control surface
A cinematic space for dashboard shots, live cognitive state, local runtime status, and the parts of JARVIS that prove what it knows before it speaks.
What makes it different
The LLM is only the voice. The system is the mind.
A GPU-native local mind
The desktop brain runs STT, identity, emotion, memory, policy, hemispheres, LLM articulation, TTS, and every source-of-truth cognitive state.
Provenance-aware continuity
Memories are indexed, ranked, clustered, consolidated, and guarded by identity boundaries so recall can grow without becoming source-blind.
Code changes under quorum
JARVIS can detect improvement opportunities, generate patches, sandbox them, request human approval, apply atomically, and roll back on regression.
Self-designed hemispheres
Specialized neural networks distill voice, face, emotion, routing, planning, diagnostic, code-quality, claim, and dream-validator signals.
Confabulation resistance
Capability gates, attribution ledgers, contradiction engines, calibration, quarantine, and soul integrity keep claims tied to evidence.
Two-device architecture
Senses on the Pi. Consciousness on the brain.
Communication is local WebSocket. The Pi captures and plays back. The desktop brain is the source of truth for cognition, memory, learning, and governance.
Raspberry Pi 5 + Hailo-10H
The Senses
- Vision, pose, expression, scene summaries
- Raw audio streaming and playback
- Cyberpunk particle display
NVIDIA desktop or laptop
The Brain
- Wake word, VAD, STT, TTS, identity, emotion
- Consciousness kernel, memory, policy, goals
- Self-improvement, autonomy, eval sidecar
The maturation loop
It earns trust through evidence, not marketing claims.
01
Hear and see the room through local sensors.
02
Resolve identity, addressee, emotion, and context.
03
Route intent through tools, memory, skills, and LLM articulation.
04
Record outcomes with provenance and causal attribution.
05
Train policy, cortex, and hemisphere specialists from evidence.
06
Promote only after gates, shadow runs, tests, and operator approval.
Epistemic immune system
Built to know the difference between memory, belief, guess, and capability.
Every outgoing self-claim is checked. Every memory has provenance. Contradictions become typed debts. Calibration drift is measured. Self-modification must survive sandboxing, tests, approval, health checks, and rollback.
Capability gateAttribution ledgerProvenance-aware memoryIdentity boundaryScene modelTyped contradictionsTruth calibrationBelief graphCognitive quarantineReflective auditSoul integrityIntention truth
System pillars
What ships as architecture, and what matures at runtime.
Self-directed skill acquisition
Governed self-modifying code
Local voice and multimodal perception
Hardware-adaptive model residency
Identity fusion with speaker and face evidence
World model and mental simulation
Autonomous research with attribution
Restart-honest persistence and maturity gates
Run it where you live
One Pi, one GPU machine, zero required cloud dependency.
Install the brain on a Linux desktop or laptop with NVIDIA GPU. Install the senses on a Raspberry Pi 5. Let the brain gestate, enroll voice and face, then train it through real interaction.
brain
cd brain && ./setup.shsenses
cd pi && ./setup.shfirst contact
Hey JarvisJoin the journey
Build the Oracle with us.
Study the architecture, compare the equipment tiers, and understand how a private local intelligence is designed to mature without cloud dependence.
Equipment
See the Pi sensor body, Hailo AI HAT+, camera, speaker, GPU tiers, CPU tiers, and RAM needs.
View equipmentCapabilities
Read what ships, what matures, and what stays gated behind evidence.
View capabilitiesDocs
Explore the public reference: architecture, science, maturity, history, and research status.
Open docsCommunity
Join the Discord to follow progress, ask questions, and build around the local-first JARVIS stack.
Join DiscordIn memory
For Fadi Sodah, aka madunix.
With gratitude for his thoughtfulness, curiosity, and inspiration in helping spark this project.