{ "_note": "SOURCE OF TRUTH for centra.krinc.in. Canonical record of every Centra project, phase, decision, and progress point. The Next.js hub renders this at build time. Claude and humans edit it here; push → Vercel redeploys. Progress % is computed from task `done` flags at render — never hardcode it. Keep valid JSON.", "meta": { "name": "Centra", "tagline": "One brain. Every project orbits it.", "domain": "centra.krinc.in", "owner": "iKrinc", "updated": "2026-08-18", "workflow": "Centra is a modular brain: a one-time core plus projects that attach as modules, added as you go. One project is active at a time; the rest wait. Centra (the core) is always live because every project runs on it. Sequencing: build the core to the Blu-enabling milestone (through S3, the event backbone), then pivot to Blu and build the first module end-to-end; return to the core (S4+) as later modules need more. New ideas join the pipeline as future modules — all coordinated through Centra.", "activeProject": "blu" }, "projects": [ { "id": "centra", "codename": "Centra", "kind": "core", "status": "active", "title": "The Brain", "tagline": "Local-first, event-driven core that every project plugs into.", "summary": "The always-on brain: local AI, a durable event bus, a self-describing device registry, and a power-cut-aware health monitor. It runs on the laptop (Intel i7 + RTX 4060) today and migrates to a dedicated NAS later — and because every service is containerized, that migration is a copy-paste rather than a rebuild. Centra is not a project you finish; it is the ground every other project stands on.", "accent": "purple", "map": { "angle": 0, "radius": 0 }, "repo": "/", "currentPhase": "S1", "tech": [ "Docker Compose", "FastAPI", "Postgres + pgvector", "TimescaleDB", "NATS JetStream", "MQTT (Mosquitto)", "Redis", "MinIO", "Ollama", "Tailscale", "Caddy", "Authentik" ], "architecture": { "summary": "A modular brain: a small always-on core set up once, plus projects that attach as manifest-declared modules you add as you go. The core owns the brain-level work — coordination, memory, reasoning, anticipation, scheduling, notification — and delegates specialized capture/actuation to modules. It does most brain things, not everything.", "layers": [ { "name": "Core — always on", "role": "one-time setup; never a per-project rebuild", "items": [ "NATS JetStream event bus + store", "Postgres (pgvector + TimescaleDB)", "Redis cache", "MinIO objects", "Device gateway (MQTT · registry · health)" ] }, { "name": "Brain — the AI", "role": "does most brain things; delegates the rest", "items": [ "Orchestrator (local LLM: Qwen3.5-9B / Qwen3-7B)", "MCP tool layer (modules expose tools)", "Memory + RAG (pgvector, evolving profiles)", "Anticipation engine (scheduled + event reasoning)", "Notifier" ] }, { "name": "Module manager", "role": "manifest-driven; add / upgrade / remove at runtime", "items": [ "Module manifest (capabilities, events, tools, data, UI, lifecycle)", "Capability registry (semantic action → connector)", "Lifecycle hooks (install / upgrade / uninstall)" ] }, { "name": "Modules — the projects", "role": "attach as you go; own their domain", "items": [ "Blu · reminders", "Aura · sensing", "Sentry · vision", "Atlas · knowledge", "new ideas → pipeline" ] } ], "flow": [ "A module registers via its manifest → capability registry, event subscriptions, and MCP tools are wired in; the core is unchanged.", "Devices and modules publish events (MQTT edge → gateway → NATS JetStream) — durable and replayable.", "The orchestrator reacts to events and, on a schedule, reasons over shared state (events, telemetry, memory) using module-provided MCP tools.", "The anticipation engine predicts likely needs and acts proactively — or asks — then notifies you.", "State is written back to the shared data plane so the next decision is better-informed." ] }, "features": { "summary": "Centra is a limited JARVIS/FRIDAY — a support system, not an autonomous do-everything agent. It owns the brain-level work (remembering, watching, reasoning, anticipating, reminding, coordinating) and keeps you in the loop for anything consequential. It assists; it does not autonomously build, spend, or send on your behalf.", "groups": [ { "name": "Anticipation & memory", "items": [ "Persistent memory of your preferences, patterns, and decisions", "Scheduled reasoning that surfaces what needs you before you ask", "Reactive reasoning on events via the local LLM + module tools" ] }, { "name": "Daily support", "items": [ "Morning / evening briefings from the live state", "Reminders & follow-through (the Blu module)", "Severity- and power-cut-aware notifications, one place" ] }, { "name": "Your world", "items": [ "Home & ambient awareness (Aura / Sentry)", "Personal knowledge RAG with citations (Atlas)", "Local voice — wake, ask, speak, fully offline" ] }, { "name": "Project & build support", "items": [ "Tracks the plan from the source of truth", "Preps checklists, scripts, and status — it does not autonomously write or ship code" ] } ], "limits": [ "No irreversible or outward action without your confirmation", "No autonomous engineering — building stays with you", "Local-first; nothing leaves the house without consent", "A small, typed, fully-logged tool surface" ] }, "phases": [ { "id": "S0", "title": "Foundation & runtime", "status": "done", "summary": "Containers, repo-as-infra, and verified GPU access — the bedrock.", "objective": "Establish a reproducible, version-controlled runtime where every service is a container defined in git, and the RTX 4060 is usable from inside containers. Everything downstream assumes this substrate.", "scope": [ "Docker Engine + Compose on the host (Ubuntu Server or WSL2)", "NVIDIA Container Toolkit for GPU passthrough", "Git repository holding all compose files and configs (infra-as-code)", "Base bridge network, named volumes, and a static LAN hostname" ], "outOfScope": [ "Any application logic or AI models (later phases)", "Remote access / public exposure (S5)" ], "approach": [ { "decision": "Containerize everything from day one", "rationale": "The whole laptop→NAS migration reduces to copying compose files and volumes. This is the single highest-leverage decision in the project." }, { "decision": "Ubuntu Server preferred; WSL2 + Docker Desktop as the pragmatic fallback", "rationale": "Native Linux gives clean GPU passthrough. If the machine stays a daily-driver Windows box, WSL2 works — but container volumes must live in the Linux filesystem." }, { "decision": "Repo is the source of infra truth", "rationale": "No snowflake config. Any service can be rebuilt from the repo alone." } ], "stack": [ "Docker", "Docker Compose", "NVIDIA Container Toolkit", "Git", "Portainer" ], "tasks": [ { "t": "Git repo + SSH remote (github-krinc → iKrinc/Centra)", "done": true }, { "t": "Containerized brain stack authored (docker-compose.yml)", "done": true }, { "t": "Gateway, edge agent, and infra configs written", "done": true }, { "t": "Install Docker Desktop (WSL2 backend)", "done": false }, { "t": "Install NVIDIA Container Toolkit", "done": false }, { "t": "Verify the RTX 4060 inside a container (nvidia-smi)", "done": false } ], "deliverables": [ "docker-compose.yml + gpu/devices overlays", "Reproducible `docker compose up` that brings the brain online", "Portainer for at-a-glance container health" ], "acceptance": [ "`docker run --rm --gpus all` reports the 4060", "`docker compose up -d` starts all core services with healthchecks green", "A fresh clone can recreate the stack with no manual steps" ], "risks": [ { "risk": "WSL2 volume I/O on /mnt/c is extremely slow", "mitigation": "Keep all named volumes inside the WSL2 Linux filesystem." }, { "risk": "GPU not visible in containers", "mitigation": "Match NVIDIA driver + Container Toolkit versions; test with the CUDA base image first." } ], "dependencies": [], "references": [ { "title": "NVIDIA Container Toolkit — install guide", "url": "https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html" }, { "title": "Docker Compose reference", "url": "https://docs.docker.com/compose/" } ] }, { "id": "S1", "title": "Local AI stack", "status": "active", "summary": "A local LLM, STT, and TTS running on the 4060 — fully offline.", "objective": "Prove the core premise: capable AI on your own hardware with no internet. This is the fastest motivating win and the model backend every project will call.", "scope": [ "Ollama serving a quantized 7–9B model on the GPU", "Open WebUI as a chat front-end", "whisper.cpp / faster-whisper for speech-to-text", "Piper for text-to-speech", "An embedding model for later RAG" ], "outOfScope": [ "Agentic tool-calling and orchestration (S4)", "Cross-device voice pipeline (Blu's project)" ], "approach": [ { "decision": "Qwen3.5-9B (or Qwen3-7B) at Q4_K_M via Ollama/GGUF", "rationale": "2026 sweet spot for 8 GB VRAM — ~40–55 tok/s fully in-GPU, strong reasoning/coding scores, and near-lossless under KV-cache quantization for long context." }, { "decision": "One large model resident at a time (OLLAMA_MAX_LOADED_MODELS=1)", "rationale": "8 GB holds a single 7–9B Q4 model well; whisper/piper/embeddings coexist since they're small or CPU-bound." }, { "decision": "whisper.cpp for STT", "rationale": "Fast, CPU/GPU flexible, no Python runtime bloat; tiny/base for realtime, larger models when accuracy matters." } ], "stack": [ "Ollama", "Open WebUI", "whisper.cpp", "Piper", "nomic-embed-text" ], "tasks": [ { "t": "Ollama container on the GPU", "done": false }, { "t": "Pull Qwen3.5-9B / Qwen3-7B (Q4_K_M)", "done": false }, { "t": "Open WebUI chat front-end", "done": false }, { "t": "whisper.cpp (STT) service", "done": false }, { "t": "Piper (TTS) service", "done": false }, { "t": "Embedding model wired for RAG", "done": false } ], "deliverables": [ "A browser chat against a local LLM", "An STT endpoint", "A TTS endpoint" ], "acceptance": [ "Chat with the local LLM in the browser with Wi-Fi off", "Transcribe a WAV file locally in near real-time", "Synthesize a spoken sentence from text locally", "Sustained tokens/sec measured and recorded" ], "risks": [ { "risk": "VRAM exhaustion / spill to CPU", "mitigation": "Stick to Q4_K_M; cap loaded models; watch KV-cache size and context length." }, { "risk": "Thermal throttling on the laptop", "mitigation": "Elevate/cool the chassis; monitor GPU temp; keep sustained load reasonable." } ], "dependencies": [ "S0" ], "references": [ { "title": "Best Local LLMs for 8GB VRAM (2026)", "url": "https://localllm.in/blog/best-local-llms-8gb-vram-2025" }, { "title": "Ollama VRAM requirements (2026)", "url": "https://localllm.in/blog/ollama-vram-requirements-for-local-llms" }, { "title": "Ollama", "url": "https://ollama.com" }, { "title": "whisper.cpp", "url": "https://github.com/ggerganov/whisper.cpp" } ] }, { "id": "S2", "title": "Data layer", "status": "todo", "summary": "One Postgres for relational + vectors + time-series, plus cache and object store — with restores that actually work.", "objective": "Give the brain a single, durable source of truth for events, embeddings, and telemetry, and prove the backup can be restored — not just written.", "scope": [ "Postgres with pgvector + TimescaleDB in one instance", "Redis for cache and sessions", "MinIO for blobs (audio clips, images)", "Automated backups (pg_dump + restic) with a tested restore" ], "outOfScope": [ "Offsite/cloud backup (S8)", "Analytics dashboards (S8)" ], "approach": [ { "decision": "timescaledb-ha image = Postgres + TimescaleDB + pgvector", "rationale": "One database engine instead of three. Telemetry is a Timescale hypertable; embeddings use pgvector; relational data is plain Postgres." }, { "decision": "Postgres is the single source of truth", "rationale": "Events land here first; the bus is a fan-out convenience. If the bus is down, no event is lost." }, { "decision": "Redis is cache/sessions only, never a bus", "rationale": "Avoid three overlapping message systems; keep responsibilities crisp." } ], "stack": [ "Postgres 16", "TimescaleDB", "pgvector", "Redis", "MinIO", "restic" ], "tasks": [ { "t": "Postgres + pgvector + TimescaleDB (one instance)", "done": true }, { "t": "Redis cache + MinIO object store", "done": true }, { "t": "Gateway auto-creates schema on startup", "done": true }, { "t": "Automated pg_dump + restic to a backup drive", "done": false }, { "t": "Perform and verify one full restore", "done": false } ], "deliverables": [ "Running data services with healthchecks", "A scheduled backup job", "A documented, verified restore procedure" ], "acceptance": [ "All data services healthy and reachable from the gateway", "A backup is produced on schedule", "A restore into a scratch database reproduces the data end-to-end" ], "risks": [ { "risk": "Silent backup that never restores", "mitigation": "Restore drills are part of the DoD, not optional." }, { "risk": "Consumer SSD wear / single disk", "mitigation": "Backups to a separate drive now; ZFS redundancy at NAS migration (S9)." } ], "dependencies": [ "S0" ], "references": [ { "title": "TimescaleDB", "url": "https://docs.timescale.com" }, { "title": "pgvector", "url": "https://github.com/pgvector/pgvector" }, { "title": "restic", "url": "https://restic.net" } ] }, { "id": "S3", "title": "Event backbone & API", "status": "todo", "summary": "The event schema and the gateway every service and device speaks to.", "objective": "Define the event contract and stand up the ingest path so devices and services communicate through one durable, replayable bus.", "scope": [ "NATS JetStream as internal bus + event store", "Mosquitto MQTT broker for the device edge", "FastAPI gateway: ingest, device registry, health monitor", "A frozen, versioned event schema" ], "outOfScope": [ "AI decisioning (S4)", "Device firmware (project-specific)" ], "approach": [ { "decision": "MQTT for the edge, NATS JetStream for internal", "rationale": "MQTT is what constrained devices speak (retain, QoS, LWT). NATS JetStream is the durable, replayable internal bus and event store. Redis stays out of messaging." }, { "decision": "Idempotent, ordered events (event_id + seq)", "rationale": "Devices draining a store-and-forward outbox after reconnect must never double-count; the gateway dedupes on event_id." }, { "decision": "Power-cut-aware health monitor", "rationale": "A mains device going silent during a grid outage is expected; a battery device going silent is a fault. Device power_class + grid state drive alerting." }, { "decision": "After S3, pivot to Blu", "rationale": "S1–S3 give Blu everything it needs — local AI, the data plane, and the event backbone. Blu is the first module; build it end-to-end before deepening the core (S4+ returns as later modules require)." } ], "stack": [ "NATS JetStream", "Mosquitto", "FastAPI", "asyncpg", "aiomqtt" ], "tasks": [ { "t": "NATS JetStream bus + event store", "done": true }, { "t": "Mosquitto MQTT broker with persistence + LWT", "done": true }, { "t": "Gateway: ingest + registry + health monitor", "done": true }, { "t": "Freeze and document the event schema", "done": false } ], "deliverables": [ "Running gateway + brokers", "Documented device contract & event schema", "HTTP API for devices/events/alerts" ], "acceptance": [ "publish → worker consumes → persists → triggers an action, end-to-end", "A duplicate event (same event_id) is deduped", "A silent battery device raises a fault; a mains device during grid-down does not" ], "risks": [ { "risk": "Schema churn breaks every consumer", "mitigation": "Freeze the core schema early; version it; additive changes only." } ], "dependencies": [ "S0", "S2" ], "references": [ { "title": "NATS JetStream", "url": "https://docs.nats.io/nats-concepts/jetstream" }, { "title": "Eclipse Mosquitto", "url": "https://mosquitto.org/documentation/" }, { "title": "MQTT Last Will & Testament", "url": "https://www.hivemq.com/blog/mqtt-essentials-part-9-last-will-and-testament/" } ] }, { "id": "S4", "title": "AI Orchestrator", "status": "todo", "summary": "The differentiated core: events in, typed tool-calls to the LLM, actions out.", "objective": "Turn raw events into intelligent actions. This is the one piece worth building custom — the glue that makes Centra more than the sum of its containers.", "scope": [ "A service subscribing to centra.events.*", "Typed tool-calling against the local LLM", "RAG over your own data via pgvector", "Action dispatch back onto the bus / to devices" ], "outOfScope": [ "Per-project skills (each project owns its own logic)" ], "approach": [ { "decision": "Build on FastAPI + a typed agent layer (PydanticAI / LangGraph)", "rationale": "Structured, testable tool-calling beats free-form prompting for reliability; the tool surface stays small and typed." }, { "decision": "Log every decision and tool call", "rationale": "When the model does something unexpected, you need the trace. Observability is part of the design, not an afterthought." } ], "stack": [ "FastAPI", "PydanticAI / LangGraph", "Ollama", "pgvector" ], "tasks": [ { "t": "Subscribe to centra.events.*", "done": false }, { "t": "Typed tool-calling on the local LLM", "done": false }, { "t": "RAG over your own data (pgvector)", "done": false }, { "t": "Decision/tool-call logging", "done": false } ], "deliverables": [ "An orchestrator service", "A small, typed tool registry", "A decision log" ], "acceptance": [ "An event ('remind me…', a sensor trip) flows through the LLM and yields a correct action", "Every tool call is logged with inputs and outputs", "RAG returns grounded answers from your own documents" ], "risks": [ { "risk": "LLM picks wrong / unsafe tools", "mitigation": "Small typed tool surface; confirmations for side-effectful actions; full logging." } ], "dependencies": [ "S1", "S3" ], "references": [ { "title": "PydanticAI", "url": "https://ai.pydantic.dev" }, { "title": "LangGraph", "url": "https://langchain-ai.github.io/langgraph/" } ] }, { "id": "S5", "title": "Secure remote access", "status": "todo", "summary": "Reach Centra from anywhere with zero open router ports.", "objective": "Make the brain reachable off-LAN securely, without exposing anything to the public internet until there's a real need.", "scope": [ "Tailscale on the brain + phone/laptop", "Caddy reverse proxy with automatic TLS on the tailnet", "Authentik/Authelia forward-auth in front of web services" ], "outOfScope": [ "Public exposure / port-forwarding (deliberately avoided)" ], "approach": [ { "decision": "Tailscale-first, public-last", "rationale": "A tailnet gives encrypted, authenticated access with no open ports. Public exposure is only added when sharing with someone who can't join the tailnet." }, { "decision": "Caddy over raw Nginx", "rationale": "Automatic HTTPS and far simpler config for the reverse-proxy + TLS need." }, { "decision": "Authentik/Authelia over Keycloak", "rationale": "SSO + 2FA at a fraction of Keycloak's weight for a personal system." } ], "stack": [ "Tailscale", "Caddy", "Authentik / Authelia", "Let's Encrypt" ], "tasks": [ { "t": "Tailscale on the brain + phone/laptop", "done": false }, { "t": "Caddy reverse proxy + TLS on the tailnet", "done": false }, { "t": "Authentik/Authelia forward-auth", "done": false } ], "deliverables": [ "A private tailnet", "TLS-terminated internal hostnames", "SSO in front of services" ], "acceptance": [ "From cellular (not home Wi-Fi), reach the dashboard securely", "Services require login via the IdP", "No inbound router ports are opened" ], "risks": [ { "risk": "Accidental public exposure", "mitigation": "No port-forwarding; audit that services bind to the tailnet/localhost only." } ], "dependencies": [ "S3" ], "references": [ { "title": "Tailscale", "url": "https://tailscale.com/kb/" }, { "title": "Caddy automatic HTTPS", "url": "https://caddyserver.com/docs/automatic-https" }, { "title": "Authentik", "url": "https://goauthentik.io" } ] }, { "id": "S6", "title": "IoT & Home Assistant", "status": "todo", "summary": "The physical world plugs into the bus via Home Assistant + ESPHome.", "objective": "Adopt Home Assistant as the IoT spine rather than hand-building device management, and bridge its events into Centra's bus.", "scope": [ "Home Assistant with MQTT", "ESP32 devices flashed via ESPHome", "A bridge from HA events into NATS" ], "outOfScope": [ "Custom automation engine (HA already does this)" ], "approach": [ { "decision": "Home Assistant + ESPHome as the IoT layer", "rationale": "1000+ integrations, device management, scenes, dashboards, local voice — building this from scratch would cost months." }, { "decision": "Bridge HA → NATS", "rationale": "The orchestrator acts on physical events through the same bus as everything else." } ], "stack": [ "Home Assistant", "ESPHome", "Zigbee2MQTT (optional)", "MQTT" ], "tasks": [ { "t": "Home Assistant + MQTT", "done": false }, { "t": "ESP32s via ESPHome", "done": false }, { "t": "Bridge HA events into NATS", "done": false } ], "deliverables": [ "A running HA instance", "At least one ESPHome device", "HA↔NATS bridge" ], "acceptance": [ "A physical sensor event reaches Centra's bus", "The orchestrator can respond to it", "Devices survive an HA restart (retained state)" ], "risks": [ { "risk": "USB/Zigbee coordinator instability on the brain", "mitigation": "Host coordinators on the Pi, physically near the devices." } ], "dependencies": [ "S3" ], "references": [ { "title": "Home Assistant", "url": "https://www.home-assistant.io" }, { "title": "ESPHome", "url": "https://esphome.io" } ] }, { "id": "S7", "title": "Clients", "status": "todo", "summary": "Your face on the system — a web dashboard and a phone client.", "objective": "Give Centra a first-class UI: live state and control from the web and your phone.", "scope": [ "Next.js dashboard against the gateway API", "A PWA before any native app" ], "outOfScope": [ "This public hub site (separate; it's the record, not the control panel)" ], "approach": [ { "decision": "PWA first, native later", "rationale": "A PWA delivers ~90% of the value for ~10% of the effort; go native only if a capability truly requires it." } ], "stack": [ "Next.js", "PWA", "Authentik (auth)" ], "tasks": [ { "t": "Next.js dashboard against the gateway API", "done": false }, { "t": "PWA packaging", "done": false } ], "deliverables": [ "A live dashboard", "An installable PWA" ], "acceptance": [ "Dashboard shows live device/event state", "You can issue a command from your phone over Tailscale" ], "risks": [ { "risk": "Scope creep into a full native app early", "mitigation": "Ship the PWA; defer native." } ], "dependencies": [ "S3", "S5" ], "references": [ { "title": "Next.js", "url": "https://nextjs.org/docs" }, { "title": "PWA on web.dev", "url": "https://web.dev/learn/pwa/" } ] }, { "id": "S8", "title": "Observability, backup & hardening", "status": "todo", "summary": "The sleep-at-night layer: metrics, logs, offsite DR, and secrets.", "objective": "Make the brain observable and recoverable, and remove plaintext secrets — so a failure is visible and reversible.", "scope": [ "Prometheus + Grafana (incl. GPU temp/util)", "Uptime Kuma + Loki logs", "Offsite backup (restic → Backblaze B2)", "Secrets → Docker secrets / SOPS" ], "outOfScope": [ "NAS hardware (S9)" ], "approach": [ { "decision": "restic to Backblaze B2 for offsite", "rationale": "Cheap (~$6/TB·mo), encrypted, S3-compatible; pairs with the local restic backups from S2." }, { "decision": "SOPS for secrets in git", "rationale": "Encrypted secrets can live in the repo safely; no plaintext .env in version control." } ], "stack": [ "Prometheus", "Grafana", "Loki", "Uptime Kuma", "restic", "Backblaze B2", "SOPS" ], "tasks": [ { "t": "Prometheus + Grafana (incl. GPU temp)", "done": false }, { "t": "Uptime Kuma + Loki logs", "done": false }, { "t": "Offsite backup (restic → Backblaze B2)", "done": false }, { "t": "Secrets → Docker secrets / SOPS", "done": false } ], "deliverables": [ "Dashboards for host/GPU/services", "Offsite encrypted backups", "Encrypted secrets in the repo" ], "acceptance": [ "Dashboards show CPU/GPU/temp and service health", "A simulated disk-loss recovery from offsite backup succeeds", "No plaintext secret exists in git history" ], "risks": [ { "risk": "Backups written but never restored", "mitigation": "DR drill is part of the DoD." } ], "dependencies": [ "S2", "S3" ], "references": [ { "title": "Grafana", "url": "https://grafana.com/docs/" }, { "title": "SOPS", "url": "https://github.com/getsops/sops" }, { "title": "Backblaze B2", "url": "https://www.backblaze.com/cloud-storage" } ] }, { "id": "S9", "title": "Expansion & NAS migration", "status": "todo", "summary": "Move the brain to dedicated hardware — the payoff for containerizing everything.", "objective": "Retire the laptop-as-server: stand up a NAS/mini-PC, migrate the stack, and free the laptop to be just a client.", "scope": [ "Build/buy a NAS (TrueNAS SCALE or Ubuntu + ZFS)", "Move compose + volumes; repoint Tailscale", "ZFS redundancy + UPS via NUT" ], "outOfScope": [], "approach": [ { "decision": "DIY/mini-PC NAS over turnkey Synology for a Docker + ZFS box", "rationale": "Better value and full control; ~$340–525 DIY 4-bay vs ~$960 Synology DS923+ (2026)." }, { "decision": "Real UPS + NUT replaces the laptop battery", "rationale": "The grid-state feed moves from the laptop's AC status to a proper UPS; graceful shutdown on low battery." } ], "stack": [ "TrueNAS SCALE / Ubuntu + ZFS", "Docker", "NUT (Network UPS Tools)" ], "tasks": [ { "t": "Build/buy NAS (TrueNAS SCALE or Ubuntu + ZFS)", "done": false }, { "t": "Move compose + volumes; repoint Tailscale", "done": false }, { "t": "ZFS pool + UPS/NUT graceful shutdown", "done": false } ], "deliverables": [ "A dedicated always-on brain", "ZFS-backed storage", "UPS-protected graceful shutdown" ], "acceptance": [ "Centra runs on the dedicated box; the laptop is just a client", "A drive-loss test is survived by ZFS redundancy", "A power cut triggers a clean shutdown via NUT" ], "risks": [ { "risk": "Migration downtime / data loss", "mitigation": "ZFS send + restic restore; cut over only after a verified copy exists." } ], "dependencies": [ "S0", "S2", "S8" ], "references": [ { "title": "TrueNAS SCALE", "url": "https://www.truenas.com/truenas-scale/" }, { "title": "Network UPS Tools", "url": "https://networkupstools.org" }, { "title": "DIY NAS vs Synology (2026)", "url": "https://homelabpicks.com/nas/diy-nas-vs-synology/" } ] } ], "log": [ { "date": "2026-08-18", "entry": "Scaffolded the containerized brain (compose, gateway, edge agent, infra) and pushed to git. Added the power-cut-aware resilience architecture." }, { "date": "2026-08-18", "entry": "Rebuilt the hub as a Next.js app (Vercel-ready, centra.krinc.in) with an animated project map, per-project overview pages, and engineered phase pages." }, { "date": "2026-08-18", "entry": "Added CLAUDE.md operating manual + per-project folder convention so the AI works directly from this repo against the source of truth (web/content/centra.json)." }, { "date": "2026-08-18", "entry": "Redesigned the home as a revolving 3D atom (crisp, no hero text; hover pauses + zooms; orbs grow with progress) and added an animated atom logo. Documented Centra's feature set as a limited support assistant (docs/centra-features.md) and shipped a one-shot S1 local-AI bring-up script." }, { "date": "2026-08-18", "entry": "S0 complete: Docker Desktop running and the RTX 4060 (driver 610.88, 8 GB) verified inside a container. Started S1 — bringing up the local AI stack (Ollama + Qwen + embeddings + Open WebUI) on the GPU. Home tuned to a clean 2D orbital rotation with 3D orbs; base page has no header/footer." } ] }, { "id": "blu", "codename": "Blu", "kind": "project", "status": "active", "title": "Pocket Voice Reminder", "tagline": "Speak a reminder into your pocket. It hears you offline; Centra remembers.", "summary": "Blu is a pocket, always-listening voice device. You mutter 'remind me to water the plants' and Blu — running fully offline — captures it, transcribes it locally, and turns it into a reminder. It never waits for the network: it responds on the spot and syncs to Centra when a link appears, where the intent is parsed, scheduled, and surfaced back to you. Blu is the first real project riding the brain, and the proving ground for the whole resilient-edge contract.", "accent": "blue", "map": { "angle": 205, "radius": 0.66 }, "repo": "/projects/blu", "currentPhase": "B0", "tech": [ "EdgeAgent (store-and-forward)", "ESP32-S3", "INMP441 (I2S mic)", "MAX98357A (I2S amp)", "microWakeWord / ESP-SR", "whisper.cpp", "MQTT", "FastAPI" ], "architecture": { "summary": "Blu is the first Centra module: an ESP32-S3 pocket device captures speech locally (offline-first) and talks to the brain over MQTT via the shared EdgeAgent; a reminder module on the brain parses intent through the orchestrator, schedules, and notifies.", "layers": [ { "name": "Device (edge)", "role": "runs and responds locally, offline-first", "items": [ "ESP32-S3 (8MB PSRAM)", "INMP441 I2S mic", "MAX98357A I2S amp", "microWakeWord / ESP-SR wake word", "whisper.cpp tiny/base" ] }, { "name": "Edge runtime", "role": "never waits for the brain", "items": [ "EdgeAgent store-and-forward outbox", "Heartbeats + MQTT Last-Will", "Battery-aware conserve mode" ] }, { "name": "Brain module", "role": "the reminder capability", "items": [ "reminder-service (ingest → Postgres)", "Intent parse via orchestrator (typed)", "Scheduler (due-time queue)", "Notifier (push / TTS / phone)" ] }, { "name": "Manifest", "role": "how Blu plugs into Centra", "items": [ "events: reminder.capture (pub) · reminder.due (sub)", "tools: create_reminder · list_reminders · snooze", "data: reminders table", "ui: reminders card" ] } ], "flow": [ "Wake word / button → record → local STT → local ack (works with no network).", "reminder.capture event → durable outbox → syncs to the brain on reconnect.", "Gateway → NATS → reminder-service stores it; orchestrator parses intent → due time.", "Scheduler fires reminder.due → notifier reaches you across channels; snooze / ack loop.", "A reminder that fired while you were offline resurfaces on reconnect — nothing is lost." ] }, "phases": [ { "id": "B0", "title": "Define & scaffold", "status": "active", "summary": "Give Blu a home, a spec, and a device runtime built on the resilient-edge contract.", "objective": "Establish Blu as a first-class project: a repo folder, a clear spec, a laptop-simulated device that registers with Centra, and a brain-side reminder service — all before any hardware is bought.", "scope": [ "projects/blu folder + README", "Blu edge agent extending the shared EdgeAgent", "Reminder service skeleton (brain-side consumer)", "docker-compose.blu.yml overlay", "Hardware target selection" ], "outOfScope": [ "Real firmware (B4)", "Wake word (B1/B5)" ], "approach": [ { "decision": "Simulate on the laptop before buying hardware", "rationale": "Get the software loop — capture → transcript → reminder → sync → notify — correct first; hardware is cheap once the loop works." }, { "decision": "Reuse the EdgeAgent store-and-forward runtime", "rationale": "Blu inherits durable offline buffering, heartbeats, Last-Will, and battery-aware conserve mode for free." } ], "stack": [ "Python", "EdgeAgent", "MQTT", "FastAPI" ], "tasks": [ { "t": "projects/blu folder + README + web sprints", "done": true }, { "t": "Blu edge agent extending EdgeAgent", "done": true }, { "t": "Reminder service skeleton (brain-side consumer)", "done": true }, { "t": "docker-compose.blu.yml overlay", "done": true }, { "t": "Pick target hardware (ESP32-S3 vs Pi Zero 2 W)", "done": false } ], "deliverables": [ "A simulated Blu device", "A reminder service", "A compose overlay to run both" ], "acceptance": [ "`blu` simulates on the laptop and registers with Centra", "Its reminder events land in the event store", "Pulling the network doesn't stop local responses; events sync on reconnect" ], "risks": [ { "risk": "Premature hardware purchase", "mitigation": "Hardware selection is the LAST task of B0; the loop must work in simulation first." } ], "dependencies": [ "Centra S3" ], "references": [ { "title": "Blu README (repo)", "url": "https://github.com/iKrinc/Centra/tree/main/projects/blu" } ] }, { "id": "B1", "title": "Local capture + STT", "status": "todo", "summary": "Hear, transcribe, and respond locally — offline-first.", "objective": "Capture speech and turn it into text on the device (or via Centra when online), buffering everything so nothing is lost offline.", "scope": [ "Audio capture (push-to-talk first, VAD later)", "whisper.cpp on-device transcription (tiny/base)", "Emit reminder.capture events into the outbox", "Offload heavy STT to Centra when connected" ], "outOfScope": [ "Always-on wake word (B5)", "Intent parsing (B2)" ], "approach": [ { "decision": "whisper.cpp tiny/base on-device, larger on Centra", "rationale": "Tiny/base fit constrained hardware for realtime; when online, stream audio to Centra's whisper for higher accuracy — best of both." }, { "decision": "Push-to-talk before voice-activity detection", "rationale": "A button is deterministic and battery-cheap; VAD and wake word add cost and come later." } ], "stack": [ "whisper.cpp", "I2S audio", "EdgeAgent outbox" ], "tasks": [ { "t": "Audio capture loop (push-to-talk)", "done": false }, { "t": "whisper.cpp on-device transcription", "done": false }, { "t": "Emit reminder.capture events to the outbox", "done": false }, { "t": "Online STT offload to Centra", "done": false } ], "deliverables": [ "A capture→transcript pipeline", "Buffered transcript events" ], "acceptance": [ "Speak with Wi-Fi off → Blu transcribes and buffers locally", "Events sync to Centra on reconnect with no loss", "Online mode uses Centra's larger whisper for accuracy" ], "risks": [ { "risk": "On-device STT too slow/inaccurate", "mitigation": "Fall back to buffered audio + Centra-side transcription when online." } ], "dependencies": [ "B0", "Centra S1" ], "references": [ { "title": "whisper.cpp", "url": "https://github.com/ggerganov/whisper.cpp" } ] }, { "id": "B2", "title": "Reminder intent + service", "status": "todo", "summary": "Turn 'remind me to X at Y' into a structured, scheduled reminder.", "objective": "Parse natural-language reminders into structured records with a due time, store them, and schedule firing.", "scope": [ "Intent parsing via Centra's orchestrator (local LLM)", "A reminders table + REST API", "A due-time scheduler" ], "outOfScope": [ "Notification delivery (B3)" ], "approach": [ { "decision": "Parse intent with the local LLM (typed output)", "rationale": "Natural-language time/task extraction is exactly what the S4 orchestrator's typed tool-calling is for; keeps Blu thin." }, { "decision": "Confirm ambiguous times back to the user", "rationale": "'tomorrow evening' is fuzzy; a quick confirmation beats a silently-wrong reminder." } ], "stack": [ "FastAPI", "Postgres", "Centra orchestrator", "APScheduler / NATS timers" ], "tasks": [ { "t": "Intent parse (local LLM via orchestrator)", "done": false }, { "t": "Reminders table + REST API", "done": false }, { "t": "Scheduler (due-time queue)", "done": false } ], "deliverables": [ "A reminders API", "A scheduler", "Structured reminders from speech" ], "acceptance": [ "A spoken reminder becomes a structured record with a due time", "Ambiguous times trigger a confirmation", "The scheduler fires at the due time" ], "risks": [ { "risk": "Messy natural-language time parsing", "mitigation": "Confirm ambiguous cases; normalize to explicit timestamps before storing." } ], "dependencies": [ "B1", "Centra S4" ], "references": [ { "title": "APScheduler", "url": "https://apscheduler.readthedocs.io" } ] }, { "id": "B3", "title": "Notify + sync", "status": "todo", "summary": "Fire reminders and confirm them across devices, losing nothing offline.", "objective": "Deliver due reminders reliably, support acknowledge/snooze, and replay anything missed during an outage.", "scope": [ "Notification fan-out (push / TTS / phone)", "Acknowledge + snooze loop", "Offline replay of missed reminders" ], "outOfScope": [ "Hardware TTS playback wiring (B4)" ], "approach": [ { "decision": "Reminders fire from Centra, with local echo on Blu", "rationale": "Centra is the reliable scheduler; Blu echoes locally (TTS/LED) so it works even if Blu was offline when it fired." }, { "decision": "Missed-while-offline reminders resurface on reconnect", "rationale": "A reminder that fired while you were away must still reach you — the whole point of the resilient design." } ], "stack": [ "MQTT", "Piper (TTS)", "Push (ntfy / web push)" ], "tasks": [ { "t": "Notification fan-out (push / TTS / phone)", "done": false }, { "t": "Acknowledge + snooze loop", "done": false }, { "t": "Offline replay of missed reminders", "done": false } ], "deliverables": [ "Multi-channel notifications", "Ack/snooze flow", "Offline replay" ], "acceptance": [ "A due reminder reaches you across at least two channels", "Snooze and done both work and persist", "A reminder that fired during an outage surfaces on reconnect" ], "risks": [ { "risk": "Duplicate notifications across channels/devices", "mitigation": "Idempotent notification IDs; ack collapses the fan-out." } ], "dependencies": [ "B2", "Centra S5" ], "references": [ { "title": "ntfy", "url": "https://docs.ntfy.sh" } ] }, { "id": "B4", "title": "Real hardware", "status": "todo", "summary": "Leave the laptop simulator; become a real pocket device.", "objective": "Bring the working software loop onto ESP32-S3 hardware with a mic, a button, a status LED, and a battery — a device you can actually pocket.", "scope": [ "ESP32-S3 + INMP441 (I2S mic) capture", "MAX98357A (I2S amp) for local TTS/beeps", "Button + status LED", "LiPo battery + charge management" ], "outOfScope": [ "Wake word (B5)", "Enclosure polish (B5)" ], "approach": [ { "decision": "ESP32-S3 (8 MB PSRAM) + INMP441 + MAX98357A", "rationale": "The proven 2026 DIY voice stack: I2S digital mic in, I2S amp out, enough PSRAM for audio buffers and small wake models." }, { "decision": "Plan I2S bus ownership up front", "rationale": "Known gotcha: mic (input) and amp (output) contend for I2S; use separate I2S peripherals or careful half-duplex handoff to avoid playback collisions." }, { "decision": "Battery device with store-and-forward", "rationale": "It must keep buffering through a power cut — that is the entire premise; the EdgeAgent's conserve mode manages low battery." } ], "stack": [ "ESP32-S3", "INMP441", "MAX98357A", "ESP-IDF / ESPHome", "LiPo + TP4056/charger" ], "tasks": [ { "t": "Flash firmware to ESP32-S3", "done": false }, { "t": "INMP441 mic capture over I2S", "done": false }, { "t": "MAX98357A playback (separate I2S / handoff)", "done": false }, { "t": "Button + status LED", "done": false }, { "t": "Battery + charge management", "done": false } ], "deliverables": [ "A working hardware prototype", "Mic-in / amp-out", "Battery-powered operation" ], "acceptance": [ "A physical Blu captures a reminder and it lands in Centra", "Local beep/TTS plays without colliding with capture", "It keeps buffering through a simulated power cut" ], "risks": [ { "risk": "I2S mic/amp bus collision", "mitigation": "Use separate I2S peripherals or strict half-duplex handoff; validate early with a loopback test." }, { "risk": "Battery life too short for always-on", "mitigation": "Push-to-talk baseline; defer always-on wake to B5 with a low-power path." } ], "dependencies": [ "B1", "B3" ], "references": [ { "title": "DIY ESP32-S3 Voice Assistant (INMP441)", "url": "https://www.instructables.com/DIY-ESP32-S3-Voice-Assistant-V02-Upgrade-for-Real-/" }, { "title": "ESP32-S3 INMP441 + MAX98357A: I2S bus ownership", "url": "https://community.home-assistant.io/t/custom-esp32-s3-voice-assistant-inmp441-max98357a-i2s-bus-ownership-and-playback-collisions/1012315" }, { "title": "Espressif ESP-SR (speech recognition)", "url": "https://github.com/espressif/esp-sr" } ] }, { "id": "B5", "title": "Wake word, battery & enclosure", "status": "todo", "summary": "Hands-free 'Hey Blu', all-day battery, and a pocketable body.", "objective": "Make Blu genuinely wearable: on-device wake word, tuned power budget, and a printed enclosure.", "scope": [ "On-device wake word ('Hey Blu') low-power path", "Battery-life tuning / conserve mode", "3D-printed enclosure" ], "outOfScope": [], "approach": [ { "decision": "On-device wake word via microWakeWord / ESP-SR WakeNet, not server openWakeWord", "rationale": "Research: openWakeWord is too heavy for an MCU and forces constant audio streaming; microWakeWord/WakeNet7 run on the ESP32-S3 (~320 KB model, <2% false-positive) so Blu only wakes and streams on the phrase — critical for battery." }, { "decision": "Wake word is the biggest battery cost — budget for it", "rationale": "Always-on listening dominates power draw; size the battery and duty-cycle around it, and keep push-to-talk as a fallback." } ], "stack": [ "microWakeWord / ESP-SR WakeNet", "LiPo", "FreeCAD / 3D print" ], "tasks": [ { "t": "Wake-word ('Hey Blu') low-power path", "done": false }, { "t": "Battery-life tuning / conserve mode", "done": false }, { "t": "3D-printed enclosure", "done": false } ], "deliverables": [ "Hands-free wake", "A full-day battery budget", "A carryable enclosure" ], "acceptance": [ "'Hey Blu' wakes the device reliably in a normal-noise room", "A full day on a charge with typical use", "Pocketable enclosure houses board, mic, battery" ], "risks": [ { "risk": "Wake-word false positives/negatives", "mitigation": "Tune threshold; train a custom 'Hey Blu' model; validate in real ambient noise." }, { "risk": "Battery can't sustain always-on", "mitigation": "Low-power wake path + conserve mode; fall back to push-to-talk." } ], "dependencies": [ "B4" ], "references": [ { "title": "microWakeWord (Kevin Ahrendt)", "url": "https://www.kevinahrendt.com/micro-wake-word" }, { "title": "OHF-Voice / micro-wake-word", "url": "https://github.com/OHF-Voice/micro-wake-word" }, { "title": "openWakeWord", "url": "https://github.com/dscripka/openWakeWord" }, { "title": "Home Assistant — wake words", "url": "https://www.home-assistant.io/voice_control/about_wake_word/" } ] } ], "log": [ { "date": "2026-08-18", "entry": "Blu is born as the first project. Folder, spec, edge agent, and reminder service scaffolded; phases B0–B5 defined with real hardware engineering." } ] }, { "id": "aura", "codename": "Aura", "kind": "project", "status": "planned", "title": "Ambient Home Sensing", "tagline": "The house quietly tells Centra how it feels.", "summary": "Temperature, air quality, light, motion, and energy sensors on ESP32 + ESPHome, streaming into TimescaleDB and visualized in Grafana. A roadmap slot — not started. Phases will be scoped, engineer-grade, when Aura is promoted to active.", "accent": "green", "map": { "angle": 320, "radius": 0.78 }, "repo": "/projects/aura", "currentPhase": "", "tech": [ "ESP32", "ESPHome", "Home Assistant", "TimescaleDB", "Grafana" ], "phases": [], "log": [] }, { "id": "sentry", "codename": "Sentry", "kind": "project", "status": "planned", "title": "Vision & Presence", "tagline": "Eyes at the edge — who's at the gate, all processed locally.", "summary": "Local camera vision (Frigate + YOLO, optionally a Coral TPU) turning frames into presence events on the bus. A roadmap slot — not started. Phases will be scoped when Sentry is promoted to active.", "accent": "purple", "map": { "angle": 90, "radius": 0.82 }, "repo": "/projects/sentry", "currentPhase": "", "tech": [ "Frigate", "YOLO", "Coral TPU", "MQTT" ], "phases": [], "log": [] }, { "id": "atlas", "codename": "Atlas", "kind": "project", "status": "planned", "title": "Personal Knowledge", "tagline": "A private mind that knows your notes, docs, and life.", "summary": "RAG over your own documents via pgvector — a private assistant that knows your stuff and cites it. A roadmap slot — not started. Phases will be scoped when Atlas is promoted to active.", "accent": "blue", "map": { "angle": 25, "radius": 0.72 }, "repo": "/projects/atlas", "currentPhase": "", "tech": [ "pgvector", "Ollama", "FastAPI", "MinIO" ], "phases": [], "log": [] } ] }