← Back to work
Shipped · Local-first

ARIA Assistant

A fully local voice agent that audits its own fairness at runtime — no cloud LLM ever touches your words.

ProblemLocal AI assistants are usually just a chatbot wired to a model — none of them check their own answers for bias.
ApproachScore every response on five fairness axes before it's spoken aloud, entirely on the user's own machine.
ResultA working voice assistant that runs offline on a single consumer GPU, with zero cloud calls.
SignificanceProof that fairness auditing can be a live, on-device property — not a one-time lab benchmark.
16Integrated tools
7.1 GBPeak VRAM
0Cloud LLM calls
5-axisRuntime audit

The Idea

Offline fairness metrics mean little if the deployed system drifts. ARIA scores every LLM response on five axes before surfacing it — calibration, faithfulness, consistency, equity, attribution — so fairness becomes a runtime property, not a one-time benchmark.

The Pipeline

Speech flows Whisper → Qwen3-8B → 5-axis audit → Piper TTS, entirely on a single consumer GPU. 16 tools give the agent real capabilities — search, files, system control — while peak usage stays at 7.1 GB VRAM and 0 cloud calls leave the machine.

Why It Matters

ARIA is the working proof that the fairness research is deployable: the same audit axes I study on paper run inline, in real time, on local hardware — privacy and accountability without a datacenter.

Methodology Pipeline

WHISPER STT QWEN3-8B 5-AXIS AUDIT PIPER TTS

What Makes This Different

Typical voice assistant
Round-trip

Your voice leaves the device, is processed on someone else's server, and the response is never checked for bias.

ARIA
Stays local

Every word is processed on-device, and every response is audited on five fairness axes before you hear it.

The tradeoff most local assistants skip: privacy and accountability, together

↗ View on GitHub