Concept

RadScribe is a voice-first field assistant for nuclear Health Physics (Radiation Protection) technicians. While conducting radiological surveys in anti-contamination suits, gloves, and PAPR respirators, technicians simply speak their readings ("General area, waist height, point 12, 45 millirem per hour") into a push-to-talk headset. RadScribe transcribes with a domain-tuned speech model (trained on RP/nuclear vocabulary, units, and isotope names), structures the data in real time, and speaks back immediate audio alerts if a reading exceeds posted administrative or procedural limits — no screen or touch required at any point in the field.

The Artifact

At the end of the survey, RadScribe auto-generates the finished deliverable: a color-coded dose-rate contour map and formatted survey report, ready for the RP supervisor to review, annotate, and sign — replacing today's hand-sketched paper diagrams and after-the-fact data entry.

Why Now

Cheap, fast agentic models (Gemini 3.7 Flash) and compact open-weight models (Qwen3.8-27B, Meta's sub-20GB Muse Glimmer) make it feasible for a small team to build and run an accurate, domain-tuned, on-prem/local voice pipeline — essential since plants require strict data-security and often no-cloud policies.

MVP

• Mobile app + ruggedized bluetooth push-to-talk headset (existing hardware, no custom device needed)
• Domain-fine-tuned speech-to-structured-data model for RP terminology and units
• Real-time audio limit-checking against a configurable procedure/limits table
• Auto-generated PDF survey map + report, exportable to the plant's existing RP database format
• Pilot target: RP staffing/contracting firms that supply technicians across multiple plants (faster sales cycle than a single utility) plus decommissioning sites and national labs

Concept

Voce is a voice-first oral examiner for higher education. Students call in (phone or browser, no app needed) and have a live, adaptive spoken conversation with an AI examiner that plays the role of a thesis committee member, language proficiency rater, or mock interviewer — asking real follow-up questions based on what the student just said, the way a human examiner would. After the call, Voce acts as judge: it scores the transcript against the department's own rubric, quotes the exact moments that justify each score, flags its confidence, and routes low-confidence or high-stakes sessions to a faculty member for a quick human check.

Why voice, why now

Oral defenses, comps, and language proficiency interviews can't be faked or replaced by typed text — the skill being tested is real-time spoken reasoning under pressure, and that's precisely the thing AI-generated writing (and watermarking efforts) can no longer certify for take-home work. LiveKit's new expressive voice mode makes the agent sound natural enough for students to take the practice seriously, and its Zero Data Retention stance gives universities the FERPA-grade privacy guarantee they require before letting a vendor near graded student speech. Claude Opus 5's long-running agent capability lets a single session hold the full arc of a 30-45 minute defense — asking probing follow-ups and then reasoning over the entire transcript to produce one coherent, evidence-backed verdict.

MVP

• Faculty/department portal: upload a rubric, course material, and 5-10 past graded oral exams as calibration examples
• Student dials in (phone number or web widget) anytime, no scheduling needed, and completes an adaptive 20-30 minute oral session
• Post-call report: rubric-by-rubric score, transcript quotes as evidence, an overall confidence rating, and specific weak areas to restudy
• Faculty dashboard showing AI-vs-human agreement rate over time per rubric criterion, so departments can see exactly how much to trust the score before using it for anything beyond practice
• Start with one wedge use case — dissertation/thesis defense rehearsal for grad schools — then expand to language department oral proficiency interviews and career center mock interviews using the same voice+judge engine