.png&w=256&q=75)
🤓 Latest Submissions
Meeting Shadow Agent
A Japanese engineer on an English delivery call may understand the request and still say yes before checking which environment, whose approval, by when. An unchecked yes can become an unintended commitment. Meeting Shadow Agent listens to the other side and, after each of their turns, suggests a reply. Before the meeting the engineer writes a memo: what they may agree to alone, what needs internal approval. The card names what is asked and whose decision it is, then writes the reply: agree inside the memo, defer outside it. The memo is the engineer's own statement, not verified company authority. Hypothesis, untested with buyers: the engineer uses it; the delivery lead at a software services firm, who absorbs the rework, buys it, at $15 per seat per month. Built on AssemblyAI Universal-3.5 Pro Realtime: tab audio streams from the browser straight to AssemblyAI; our server never receives audio. Each finished turn and the memo go to Gemini 2.5 Flash-Lite. A reply citing an utterance id nobody spoke is refused; whether a cited utterance supports the reply is not checked. Measured, with conditions: - 20 scripted audio cases: 58 of 62 turns returned a suggestion, median 2.02 s speech end to suggestion generated, in-process, excluding HTTP and rendering. 4 of 20 final cards agreed to something the memo withheld or the speaker left undefined. - Decision label, 3 repeats per case, text only: memo grants the request, 15 of 15 labelled it the engineer's; memo withholds it, 13 of 15 labelled it as needing approval (1 unclear, 1 provider error). - One text run: under a memo that withheld staging, 3 of 5 decisions flipped (1 no card, 1 unchanged). Limits: the builder wrote the cases. After a refusal, a changed question can still get the old deferral. No engineer has used it on a live call. Replay, no sign-in: https://tokimwc.github.io/meeting-shadow-agent/replay/ The video is a simulated call: LLM client, TTS voices, screen re-rendered from the run.
30 Sep 2026
