Yadu | Building reliable AI agents

Full-stack engineer building AI voice agents. TypeScript · React · Node · Postgres. Learning in public. Open to remote roles.

[Daanish Khazi @bertgodel]

We are releasing Ultramarine 1 Flash, a frontier-class model built for knowledge work. It is open-source and achieves similar performance to Fable 5, while being 60x cheaper.

It was post-trained by a team of 5 for ~$300k all-in. Available on API now, with weights releasing in 2 weeks.

[Yadu | Building reliable AI agents @yadukrishnapb10]

okay this is actually big

[Dheemanth Reddy @Dheemanthredy]

We’re releasing the world’s best Indic voice AI model, fully open-weight. 🇮🇳🇮🇳

With over 1.5 billion people and 22 official languages, the future of AI for India is voice-first.

Maya-2 officially outperforms ElevenLabs v3, Cartesia Sonic 3.5, and Bulbul v3 across Indian languages.

It will be completely free for every builder very soon.

[Yadu | Building reliable AI agents @yadukrishnapb10]

think of langchain like building a team of AI workers 👇

you ask an AI to research a competitor.

instead of making the LLM do everything alone, you give it tools:

🔍 one tool searches the web
📄 another reads documents
🧠 the LLM reasons over the results
✍️ another step turns everything into a report

langchain helps connect these pieces so they work together.

now imagine 100 users doing this simultaneously.

each user has their own conversation history and context. your app manages that state, while your backend handles concurrent requests.

the analogy:

LLM = brain
Tools = hands
LangChain = the wiring connecting them
Your backend = the system handling multiple users

one important distinction: langchain doesn’t automatically make everything multithreaded. concurrency, conversation state, and execution flow need to be handled by the appropriate parts of your application.

that’s the difference between calling an LLM and building a real AI application. 🚀

[Yadu | Building reliable AI agents @yadukrishnapb10]

langgraph explained without the headache:

imagine your ai agent handling 3 users at the same time.

user A: debugging code
user B: researching a topic
user C: writing a tweet

each conversation has its own state, memory, and execution flow.

that’s where langgraph + threads come in.

🧵 a thread keeps track of a conversation.
⚡ each thread gets its own state.
🔄 the graph decides what happens next.

same agent. multiple conversations. no mixing up who’s doing what.

basically, threads = separate workspaces for your ai agent.

once you get this, langgraph starts making way more sense.

[Yadu | Building reliable AI agents @yadukrishnapb10]

The hardest part of a voice agent isn’t the LLM.

It’s the ~800ms between someone finishing a sentence and your agent replying. Miss that, and it feels robotic.

Here’s where those milliseconds go, and how to cut them 🧵

[Yadu | Building reliable AI agents @yadukrishnapb10]

Took a look at the early Gemini 4 Argon reviews, and the interesting part isn’t the benchmarks 👀

It’s how well it handles long, multi-step coding tasks without constant hand-holding.

But early results are mixed, and I’d want to test it on a real production codebase before calling it a Claude killer.

Benchmarks are cool. Shipping working code is the real test. 🧑‍💻

#Gemini #AI #Coding #LLM #AIEngineering

[Yadu | Building reliable AI agents @yadukrishnapb10]

Imagine calling Amazon support, saying your order never arrived, and the AI checks the tracking, verifies the issue, and sends a replacement before you hang up.

that’s the kind of voice AI future worth watching 👀

[Yadu | Building reliable AI agents @yadukrishnapb10]

voice agents that can talk, use tools, and resolve issues mid-call.

the era of “press 1 to suffer” might actually be ending

[Yadu | Building reliable AI agents @yadukrishnapb10]

@CobusGreylingZA The gap between automating workflows and a system that can actually improve itself is fascinating. Feels like the quality of the feedback loop is going to matter just as much as the model itself.

[Yadu | Building reliable AI agents @yadukrishnapb10]

Hit a real bug in llama-cpp-python while building a CPU-only voice agent: enabling speculative decoding (draft_model=...) silently needs logits_all=True too, or the scores buffer comes out the wrong size and generation crashes past ~512 tokens.

Traced it to one line, added a regression test, PR's up:
https://t.co/NOmGr8srFy

Built while making potato-agent — an agentic voice LLM that runs with no GPU:
https://t.co/hZBSCnApeI

[Yadu | Building reliable AI agents @yadukrishnapb10]

The next big challenge in voice AI isn’t making models smarter.

It’s making conversations feel instant.

Nobody cares how intelligent your AI is if there’s an awkward 2-second pause after every sentence.

Latency is a product problem, not just a model problem.

And I think that’s where a lot of the real engineering work is happening.

[Yadu | Building reliable AI agents @yadukrishnapb10]

Voice AI in India just got more interesting.

Microsoft’s October Speech update brings an improved multilingual transcription model for 13 Indic language locales, with support for mixing languages across a conversation.

This matters because real conversations don’t stay in one language.

Building voice agents that work outside perfect English is a much more interesting engineering problem. 🇮🇳

[Yadu | Building reliable AI agents @yadukrishnapb10]

Builders: which stage is slowest in your agent? 👇

[Yadu | Building reliable AI agents @yadukrishnapb10]

The hardest part of a voice agent isn’t the LLM.

It’s the ~800ms between someone finishing a sentence and your agent replying. Miss that, and it feels robotic.

Here’s where those milliseconds go, and how to cut them 🧵

[Yadu | Building reliable AI agents @yadukrishnapb10]

The big lesson: everything has to stream.

A voice agent isn’t request → response. It’s a pipeline where every stage starts before the previous one finishes.

I’m a full-stack TS engineer learning voice AI in public. Next: I’ll measure each stage in a real agent and share the numbers.