Aditya Mishra

AI Engineer | Software Architect | Liverpool FC #YNWA

[Aditya Mishra @adi_myth]

Marking this as a moment convincing @swyx to bring @aiDotEngineer to India next year with @sanjeed_i @udayan_w

Exciting times!! 🥳

[Aditya Mishra @adi_myth]

@FishAudio DRAMA

[Aditya Mishra @adi_myth]

Introducing Indic Smart Turn, an open turn detection model for English and 11 Indian languages.

At every pause, a voice agent has to decide whether you've finished speaking. @pipecat_ai's Smart Turn v3.2 makes that call by listening to your last few seconds of speech, instead of just timing the silence. It's trained on English, Hindi, Marathi and Bengali, but has no training data for Tamil, Telugu, Kannada and other Indian languages.

So I trained a turn detection model on real phone conversations in 11 Indian languages, from @ai4bharat's IndicVoices. On a public test set of 20,431 clips, it scores 89.8% accuracy vs 81.9% for Smart Turn v3.2. On the eight languages v3.2 doesn't cover, it's 12 to 22 points ahead.

It's a one-line swap in Pipecat. The models, the 50K-clip dataset and all the code are open.

Write-up: https://t.co/OWGPB7iEI4
Model: https://t.co/ixk3I3vRy3
Code: https://t.co/ZqKH9bPNxi

[Aditya Mishra @adi_myth]

@dataBiryani @SarvamAI v3 was very good as well on our internal benchmarks but v4 is a lot better

[Aditya Mishra @adi_myth]

@pratos_ @swyx Share your playlist. Need new hype songs 🤣

[Aditya Mishra @adi_myth]

You can build your own Jev-style System 1 model.

Score the answers your app defines with a small open model, then calibrate. This works for any bounded decision: routing, approvals, agent tool checks.

I built one called Compass. Open weights, Jev-compatible API and runs locally.

287ms median on a 4090. Scores 100% easy, 87.5% standard, 60.4% hard on JevBench public tiers

Code: https://t.co/CzCTu05bae
Essay: https://t.co/QG9gmoFNgL

[Aditya Mishra @adi_myth]

@timus__ I am experimenting with using Jev for guardrails as well as online evaluations for agent traces

[Aditya Mishra @adi_myth]

@tonygentilcore Takeaways: Jev is an excellent zero-shot classifier when you have fixed options to choose from. But in case if you have high volume of data & are already labelled, fine tuning a classifier gives the biggest win. For anything else, use generative LLMs

[Aditya Mishra @adi_myth]

@pratos_ What have you done @pratos_

You have started a mini-war 🤣

[Aditya Mishra @adi_myth]

@askalphaxiv Thinking about it, it's pretty obvious. Cheapest & fastest methods first then raise to costlier & slower options

[Aditya Mishra @adi_myth]

Really cool demo.

Inspired me to build Pulse: a fully local take on real-time transcription and sentence-level sentiment analysis.
Everything runs on a Mac.

Demo on the same audio :)

Code: https://t.co/7jAw02ZSwm

[ElevenLabs Developers @ElevenLabsDevs]

Realtime sentiment analysis with Jev and ElevenLabs.

Each phrase takes on the color of the emotion it carries while the caller is still talking. Six meters on the right track the mood of the call.

[Aditya Mishra @adi_myth]

@pratos_ @convex The All Gas Hackathon?

[Aditya Mishra @adi_myth]

@manthanguptaa @IndraVahan This is the way!

[Aditya Mishra @adi_myth]

@TheCoderShow @claudeai @AnthropicAI @ClaudeDevs Missed this & couldn’t sign up. Can you help?

[Aditya Mishra @adi_myth]

@PetrBrzek This looks fun

[Aditya Mishra @adi_myth]

@pratos_ About?

[Aditya Mishra @adi_myth]

@victormustar Boeing bench test incoming soon?

[Aditya Mishra @adi_myth]

@janaksunil Surprised to see GPT 5.6 Sol with such low scores that too with Codex Harness

[Aditya Mishra @adi_myth]

@invideoOfficial @_anshulk I find Invideo’s launch/release videos one of the best! The creative production team deserves a raise!

[Aditya Mishra @adi_myth]

That’s a good distinction. A hard security constraint shouldn’t have to win a semantic search. The host should check and enforce those rules separately, then use memory retrieval for background that might help.
In the experiment, I put both through the systems’ normal memory paths, which is the boundary you’re pointing out.

I wasn’t testing these products as policy engines, so the result doesn’t mean they’re unsafe. A separate policy path also leaves the other finding untouched: on turns that needed no memory, all five returned as many results as their limits allowed.

[Aditya Mishra @adi_myth]

Most agent memory systems retrieve by ranking stored information against the current query.

I wanted to examine the decision around that: does this turn need memory at all, and which memories would actually improve the response?

I tested Mem0, LangMem, Graphiti, Letta and Amazon Bedrock AgentCore Memory on the same conversation history and background facts.

All five retrieved memories for turns that needed no context. Yet none surfaced a security rule when it would have changed the answer.

I also explored changing facts, retrieval thresholds, reranking, and the trade-off between recall and unnecessary context.

https://t.co/XMkhvsDV1k

Experiments & code:
https://t.co/CgAMe53m3B