Noman Khan

Backend dev building real things.
Voice agents · Automation · Open source.
NestJS · Python · PostgreSQL

[Noman Khan @nomcodes]

99% of apps don’t need AI. They need a good database and clean UI.
Stop chasing the trend.

[Noman Khan @nomcodes]

@xzx_slipknot Bopara, Callum Ferguson, Fernando , Benn

[Noman Khan @nomcodes]

One thing I learned the hard way is the moment agent needs to juggle booking + FAQs + upsells in 1 prompt, it gets flaky so it's better to
split responsibilities rather than trying to make it a 1 god-agent. Keep the scope of each agent strictly limited to get the most out of it

[Noman Khan @nomcodes]

Bro had 5 years to manifest this just for Ferran Torres to ruin his entire legacy in extra time.
#VamosEspaña

[dilemma @actuallyimthe]

Argentina just beat Spain at the 2026 World Cup final, 3-2.

[Noman Khan @nomcodes]

@MrBeast Bro really thought I had nothing better to do than counting his floor 😭😭

[Noman Khan @nomcodes]

@MaheshCodesX roadmap

[Noman Khan @nomcodes]

@joeyousss Yup! 💯

[Noman Khan @nomcodes]

Voice agents are a different beast once you are past the demo stage state tracking, off-topic detection, human handoff, latency, overlapping speech and all the stuff that doesn't show up in the pitch. It is a very fun problem space though.

[Jyoti Bisht @joeyousss]

Why is everyone joining VoiceAI companies?

[Noman Khan @nomcodes]

@tricalt Interesting 😎

[Noman Khan @nomcodes]

@jcubic Thanks for sharing 👍 noted

[Noman Khan @nomcodes]

@Dhanush_Nehru Congratulations bro! Keep rocking 👍

[Noman Khan @nomcodes]

@Hiteshdotcom @ankitpanditdev @piyushgarg_dev 😂😂

[Noman Khan @nomcodes]

Its real value is to handle the repetitive calls so your team isn't stuck answering the same 10 questions all day and can spend time on things that actually need a human.

Skip this conversation and the project feels like a failure even when it's actually working fine.

2/2

[Noman Khan @nomcodes]

Something I now tell every client upfront when building a voice agent:
"This is not a human replacement. It's a triage system".
People benchmark it against a human and get disappointed it's not perfect.
1/2

[Noman Khan @nomcodes]

@tricalt I am curious whether Cognee is a good fit for real-time voice agents. How is it for improving long-term conversation memory without adding noticeable latency?

[Noman Khan @nomcodes]

Agreed. LLMs are great at generating outputs but building reliable agentic systems still requires solid engineering around them. Curious to see how much of that changes over the next year

[Jared Friedman @snowmaker]

A funny thing is that LLMs are not particularly good building agentic systems. For the moment, you need human expertise to do it well.

[Noman Khan @nomcodes]

@harshul_lodha That's interesting! I would love to get a quick overview of your approach especially how you are combining voice isolation with semantic VAD.

[Noman Khan @nomcodes]

I think the hardest part of building voice agents isn't the AI but it is us.

We don't talk like books. We say "um," "uh," "wait, actually". Standard speech-to-text metrics often struggle with these "disfluencies," which leads to hallucinations or agents getting lost in the noise

[Noman Khan @nomcodes]

If you are interested in learning about it more here is an article that I found quite useful

https://t.co/4xiTyMgCfg

[Noman Khan @nomcodes]

If you are still using raw WER to measure your agent's success, then U are measuring the wrong thing. Look into Semantic Word Error Rate (SWER) or LLM-based evals.Focus on whether the agent understood the user, not just if it heard them.