firela-pa-pc

firela-pa PC — a Sovereign Personal AI FIRE advisor on your own device

firela-pa is a Sovereign Personal AI — a FIRE financial advisor that lives on your own device: your device decides what ever leaves, its memory of you never leaves home, and the cloud is used only when needed — with the question redacted, the provider swappable, and your identity never on the cloud side.

Sovereign Personal AI = personal service × three user-held rights × zero-tracking foundation - Classification right — a local, fine-tuned router on your device decides what leaves, not an OS vendor - Memory right — voice and agent memory never leave the device; transaction-level ledger data never goes to the cloud (aggregates only, under a field-level egress policy) - Cloud-leg right — anything that goes to the cloud passes a redaction gateway first, and the provider is swappable at any time

What this is

The PC distribution of firela-pa: a task-fine-tuned Qwen3-1.7B privacy router (five intents: ledger query / portfolio / market quote / local chat / redacted-cloud), shipped as GGUF (Q4_K_M default, Q8_0 quality tier; 200-question eval G4 98.0) — the same weights family as the on-board W8A8 build.

v1.0 capabilities (2026-09-24 stable)

• Five-branch local routing — what leaves the device is decided by a router on your device, not by an OS vendor. Chat, ledger, portfolio and market questions stay local or go through your own token; only advice-type questions go to the cloud — redacted first.
• Redaction gateway — CN ID/phone/address/person names, SSN, US phone, Japan My Number and email are scrubbed from the question before any cloud call; the provider is swappable.
• Local agent memory — FIRE assumptions and goals in a local SQLite store (0600), saved only on explicit "remember this", never leaving your device.
• FIRE simulation engine — real-return compounding, withdrawal-rate sensitivity sweep, and historical stress-sequence replay (US stock/bond rolling windows 1928–2025, worst-start 1966, window survival rates) — in the free tier, not behind a paid SaaS.
• Proactive audit — overspend / large-expense / FIRE-milestone checks over your own ledger, rendered locally.
• Local-only telemetry — a local SQLite diary (0600) for your own instrument panel. It never uploads anything.

Install (macOS / Linux / WSL2)

bash <(curl -fsSL https://huggingface.co/firela-ai/firela-pa-pc/resolve/main/install.sh)

Automatically: installs/starts Ollama (version gate ≥0.34) → downloads the app and router model from this repo (sha256-verified) → registers firela-router → pulls the generation model qwen2.5:3b-instruct (--no-gen to skip) → interactive 0600 config (vlt / relay credentials, may be left blank) → Time Machine exclusion → firela-pa command → smoke test. Unattended clean-machine install measured at ~90s.

Privacy boundary: inference never leaves home — routing and chat generation are fully local; ledger queries go through your own vlt token (zero-tracking foundation: the cloud side sees data, never you); cloud-bound questions are redacted first and carry only aggregate ledger data (field-level egress policy — transaction details never go to the cloud).

Source code (Apache-2.0, v1.0 snapshot + single-egress CI gate): https://github.com/firela-ai/firela-pa

License & attribution

Everything in this repo — pc-app.tar.gz, install.sh, router-merged-Q4_K_M.gguf, router-merged-Q8_0.gguf — is provided under Apache License 2.0.

The router model is a LoRA fine-tune merge derivative of Qwen3-1.7B (© 2024 Alibaba Cloud / Qwen team, Apache-2.0). Redistribution must include the LICENSE copy as per Apache-2.0 §4.

firela-pa is an AI tool, not a licensed financial advisor. Simulation results are historical scenario analyses, not predictions.