Archit Sakri

20yo dropout • Building private AI Infrastructure @tryopencomputer

[Archit Sakri @thearchitsakri]

OpenSlide is live. describe a deck. watch it build.

fully open source. self-host it. no paywalls ever.

try it today

link below 👇

[Archit Sakri @thearchitsakri]

@suhelbuilds what happens when a handoff is needed but nobody picks up? i'd make that fallback explicit too: capture the unresolved question, offer a callback only if someone owns it, and avoid confirming anything the agent couldn't verify.

[Archit Sakri @thearchitsakri]

@wasim_madha curious how 3.1 Pro handles Hindi-English switches within a sentence, especially names and numbers. do you evaluate that separately from the single-language sets?

[Archit Sakri @thearchitsakri]

@MiclauMarius for the grounding tiers, do price and booking-status claims have to pass individually? a 0.9 overall score could still let one wrong price through if the other nine claims are correct.

[Archit Sakri @thearchitsakri]

@Coffee_and_NLP @telnyx @mlopscommunity for the “one durable agent per caller” idea on your slide, how do you separate caller identity from phone number? curious how you keep histories separate when a family or office shares one number.

[Archit Sakri @thearchitsakri]

@builditfarhan i'd include cancellation in that pipeline too. when the caller interrupts, stopping generation isn't enough if audio is still queued for playback. does your setup flush that queue immediately?

[Archit Sakri @thearchitsakri]

@Rasmic one test i'd try: interrupt it from across the room while it's speaking loudly. does it still pick up your voice without transcribing its own reply? curious how you're handling echo cancellation.

[Archit Sakri @thearchitsakri]

@bigvalleyblue a useful next test could be the turn after a joke misses: "not funny, i'm serious." can the model adjust without getting stiff? that would test how it responds to feedback over a conversation, beyond a strong one-turn score.

[Archit Sakri @thearchitsakri]

@livekit @rimelabs what happens if the decision model times out mid-call? i'd want the buffered sentence held back and a safe fallback, rather than letting unchecked audio through to hit a latency target. is that fallback built into the integration?

[Archit Sakri @thearchitsakri]

@Hazem_Hegazy1 for complex calls, i'd score correct escalation separately from task completion. otherwise an agent that guesses a policy detail can look better than one that knows when to hand off. how are you measuring that in the new model?

[Archit Sakri @thearchitsakri]

@BinaryCoder @Speechmatics @roark_ai on the self-improving agents point: does each proposed change face a held-out set of calls before deployment? i'd want to catch a fix that helps the latest failure but regresses other accents or noisy callers.

[Archit Sakri @thearchitsakri]

@foin0ir i'd add an escape-to-cancel key too. letting go of the mic shouldn't force-submit half a thought, especially if the agent can run commands.

[Archit Sakri @thearchitsakri]

@tonywu120853 @kundan2510 i'd test paired clips with identical words but different delivery, then score whether the response changes appropriately. recognizing frustration is only half the test; does the agent actually give a more useful response?

[Archit Sakri @thearchitsakri]

@samgutentag @nakshatraavijay having the answer ready early makes sense. i’d keep tool execution behind a separate gate though: “yes, but make it Friday” should invalidate the first plan before anything gets booked. does Jev expose provisional decisions the runtime can cancel?

[Archit Sakri @thearchitsakri]

@SoumilBhandari for a silence-only endpoint, i'd try 300/500/800 ms as a test sweep. include mid-sentence pauses and spelled-out emails, then compare false endings with p95 delay. a semantic turn detector can help with unfinished thoughts, but i'd test it on those same calls.

[Archit Sakri @thearchitsakri]

@Gustav_Hugod did the input transcript switch languages too, or just the spoken reply? curious whether this came from language detection or the response model. very different things to debug.

[Archit Sakri @thearchitsakri]

@kritinv07 the tricky test is when the wrong letter still returns a real customer. lookup success can hide the transcription error. i'd include two valid, sound-alike addresses and check that the agent confirms which one the caller meant before using either record.

[Archit Sakri @thearchitsakri]

@SoumilBhandari i'd timestamp the caller's last sound, end-of-turn decision, first LLM token and first audible reply. TTFT alone misses endpointing and audio buffering. compare p50 and p95 over real calls; the bottleneck can change between quiet speech and hesitations.

[Archit Sakri @thearchitsakri]

@Thealfredalv are you showing a rolling translation or waiting for each sentence to finish? curious which tradeoff she preferred: faster text that can change, or a little delay with a more stable translation.

[Archit Sakri @thearchitsakri]

@speko_ai with the top four this close, how stable is the ordering across speakers? a per-speaker breakdown would help distinguish a broad improvement from a few unusually hard conversations.

[Archit Sakri @thearchitsakri]

@shreyaksingh i'd have the first call return a concrete plan to approve: attendees, room, permissions, all of it. then commit that exact plan. if anything changes after sign-off, ask again. fewer API calls is useful, but the approval has to bind to what actually happens.