Case study · October 2026 · team project

Drigil

Smart glasses that coach people with intellectual and developmental disabilities through a new job, in their supervisor's own voice.

2nd place · Anthropic's Claude for Public Benefit track · Cal Hacks 13.0 36 hours 3-person team
Team
Sarath Sundar Harisaravanan, Shraddha Khanapur, Michael Kim
Where
Cal Hacks 13.0, Palace of Fine Arts, San Francisco
Stack
Claude, Cartesia, SpacetimeDB, Moss, Factory Droid, Notability
Links
Devpost ↗

The problem

Starting a new job means learning a sequence of physical steps, often by watching someone once and then trying alone. For people with intellectual and developmental disabilities, that first stretch can mean a supervisor nearby, repeating the same guidance until the steps stick.

We wanted the supervisor to teach a task once and have that coaching stay with the worker: in the supervisor's voice, checking each step as it happens, and stepping back as the worker gets it.

What we built

Drigil has three loops: teaching a task once, coaching it live, and fixing steps that keep going wrong, with a person approving every change.

Teach once

A supervisor narrates a single demonstration. Drigil picks keyframes at the moments the hands pause, aligns them with a word-level transcript from Cartesia ink-whisper, and hands both to Claude, which writes a structured task specification.

Coach live

During the shift, Claude runs confidence-gated vision checks against the camera feed. Coaching plays in a cloned version of the supervisor's voice through Cartesia, with audio cached per line. A mastery model reduces prompting as the worker gets up to speed.

Fix what breaks

If a step keeps causing difficulty, Factory Droid starts a sandboxed agent that rewrites that step's instructions and corrections. The supervisor sees a before-and-after comparison and has to approve it before it replaces the step in the live workflow.

Key decisions

A person approves every rewrite

The remediation agent can propose a better version of a step, but it can't ship one. The supervisor reviews a before-and-after comparison first.

Sarath: in one or two sentences, why did the team insist on this, and what did it cost you (speed, complexity)?

Confidence-gated vision checks

Vision checks are confidence-gated: the system acts on a check only when it clears a confidence bar.

Sarath: what did the system do when confidence was low? Was there a trade-off you tuned during the hackathon?

Cached coaching audio

Each coaching line in the supervisor's voice is generated once and cached.

Sarath: what made caching necessary: latency on the glasses, cost, or both?

Outcome