Skip to main content
InterpScout is a teaching agent for interpreter training. It closes the full loop of an interpreting classroom — a teacher prepares course material, students practice against it, machines score every attempt, and the results flow back to the teacher — with autonomous agents performing the work at each step. It is not a demo. InterpScout runs in production at interpscout.org, used by a practicing interpreting teacher and her students. Every course, briefing, and score you can see with the test accounts was produced by the live system.
评委您好 — 如果您想先用中文快速了解产品逻辑与测试路径,请直接阅读 评委导览(中文)

The loop

missed terms
Teacher
a source, or a meeting topic
Producer agent
aligns & curates courses
Briefing agent
researches & writes briefings
Courses & Briefings
parallel-reading + pre-meeting
Students practice
sight interpretation, aloud
AI scoring
accuracy · terms · fluency
Flashcards
spaced repetition (FSRS)
Dashboards
progress & class-wide trends
Terminology base
shared across every course
Terracotta marks the two autonomous agents; gold marks the shared terminology base that every course and briefing both feeds and draws from. The loop closes back at the teacher, who sees only finished artifacts — or a clear explanation of what a source could not support. An interpreting classroom has three chronic bottlenecks: preparing bilingual teaching material is slow, pre-meeting preparation is unteachable at scale, and practice feedback depends entirely on teacher hours. InterpScout assigns an agent to each.

Producer — course production

The teacher points at a bilingual source (or picks from daily auto-discovered candidates). The Producer agent aligns it paragraph by paragraph, curates terminology across the whole document, and publishes a parallel-reading course — typically in minutes.

Briefing — pre-meeting preparation

Given a meeting topic, the Briefing agent researches the web, cross-checks sources, and delivers a 9,000-character pre-meeting knowledge briefing: domain crash course, speaker and institution profiles, contested issues, and a graded bilingual term table.

Practice — sight interpretation

Students interpret real sentences from real courses, out loud. A multimodal model listens to the audio directly and scores accuracy, terminology, and fluency — then routes every missed term into that student’s flashcard queue.

Classroom — management & analytics

Teachers create classes with invite codes, watch per-student averages, and see the terms the whole class keeps missing — data deciding what next week’s lesson should be.
What connects them is a cross-document terminology base: every course a teacher produces deposits its curated terms — with officially-graded translations and sources — into a shared, searchable glossary that gets thicker with use.

Why agents, not a chatbot

A chatbot answers questions; a classroom needs artifacts: a finished course, a finished briefing, a finished score. InterpScout’s agents run as durable, long-lived sessions that survive deployment restarts, report progress in real time, and are held to mechanical quality gates — a course cannot be saved with misaligned paragraphs, a briefing cannot be saved without its term table and verified further-reading links. When a gate rejects, the agent reads the reason and revises inside the same session. The teacher sees only the finished artifact, or a clear, specific explanation of what a source could not support. This philosophy — machines check what machines produce, and prompts are never trusted as guarantees — runs through the whole system. Read Truth Layer & Mechanical Invariants for how deep it goes.

In numbers

Real courses produced33 (e.g. 2026 Government Work Report: 218 terms, 246 aligned sentence pairs)
Pre-meeting briefings9,000-character class, with graded bilingual term tables
Practice loop3-dimension multimodal audio scoring → automatic weak-term flashcards
DeploymentProduction on Vercel, durable agent runtime, used by real teacher & students

Where to go next