Stead - Assistive Care Embodiment - built on BolospotTop 5Deep-reviewed
by Alison Cossette
What they're building
Sovereign, permission-aware voice agents for assistive robotics. A young woman with cerebral palsy runs her own life through an agent that understands her — a model fine-tuned on her own voice (Nebius H100), acting on her behalf only inside scoped, revocable, owner-held consent (Bolo) and acting on real apps (Composio). Plus Plumbline (github.com/alisoncossette/plumbline) — an open-source intent-boundary analyzer that hardens any agent acting on someone's behalf (it found & we fixed real bugs in ours).
AI code reviewrepo: real
Real, deployed product: the Cloud Run app loads a working consent dashboard (roles, ledger, calendar, Ask Stead), backed by a genuine, tested consent-engine agent (consent.act gates a real Vapi phone-order, halts/escalates over cap). Tavily (web_search) and Composio (GOOGLECALENDAR_CREATE_EVENT, GMAIL_SEND_EMAIL) are real httpx/SDK adapters genuinely wired into the live agent's tool_runner in app/server.py:319/380-383 — two load-bearing sponsors. However the headline pitch ('fine-tuned on her own voice, Nebius H100, loss 2.12->0.53') is NOT backed by the repo: model/artifacts/train.jsonl is 0 bytes, finetune_plan.json is an un-submitted plan with a placeholder subnet and a Windows path, and eval.py's Token Factory path is an explicit STUB — so technical_depth is capped per the unbacked-claim floor. ⚑ Nebius fine-tune is plan/claim only (no training data, no checkpoint, no run logs); the central 'sovereign fine-tuned model' narrative is unverifiable from the repo. Composio/Tavily live only when keys are set (offline-degrading by default).