Backend-Focused Full-Stack Developer | GenAI
Based in India
- Tweets 6,292
- Following 707
- Followers 709
- Likes 21,464
Grateful to have attended the AI Agent Masterclass by @surajtwt_
Happy to finish Top 2 out of 263 participants in the quiz
Learned more about LangGraph, agent workflows tools and state
Thankful for the opportunity to learn and test my understanding🙌
@Hiteshdotcom @nirudhuuu
3
1
30
406
I started looking beyond the LLM call.
How do we design the backend? Handle tools, workflows, memory, and failures?
Still a lot to learn, but my direction is clearer.
Grateful to @Hiteshdotcom @piyushgarg_dev @nirudhuuu .
Certificate earned. The real building starts now. 🎓
1
13
123
To Tejas:
Scan → Authenticate → Pay → ✓
To the systems underneath:
Routing → Authorization → Debit → Credit → Confirmation → Failure handling → Reconciliation
24.51 billion transactions look simple from the outside.
The engineering underneath is anything but.
3
34
Now break it.
Tejas's account gets debited.
But the confirmation never reaches his app.
He sees: "Something went wrong."
Did it fail? Should he retry? What if it actually succeeded and a retry double-charges him?
This is where it gets interesting.
1
2
27
At a simplified level, UPI involves:
Tejas's App → Payer PSP
Kunal's App → Payee PSP
Tejas's Bank ↔ NPCI / UPI ↔ Kunal's Bank
NPCI operates the UPI switch/network. The banks handle the actual debit and credit.
Let's follow Tejas's ₹500. 👇
1
2
74
Today we built a Claude Code-style CLI using the Claude SDK.
The interesting part wasn't just connecting the model.
We focused on building the experience around it tool execution, readable output, transparency, and different modes.
@Hiteshdotcom @piyushgarg_dev @nirudhuuu
1
32
427
Today’s GenAI class was about building an AI Pitch Deck generator.
The interesting part wasn’t just generating images.
We focused on how to make the AI output structured, validated, and usable by the application.
@Hiteshdotcom @surajtwt_ @nirudhuuu @ChaiCodeHQ
1
32
325
Your API request doesn’t travel alone.
It gets wrapped, addressed, forwarded, unwrapped… before your server ever sees it.
Here’s what’s actually happening behind your GET /profile.
@wjshubham
2
25
504
Today's GenAI class was about building an AutoWiki for Git repositories a RAG system for asking questions about a GitHub codebase.
But my main takeaway wasn't the UI.
It was understanding the backend architecture behind repository-aware AI.
🧵👇
@Hiteshdotcom @piyushgarg_dev
1
33
421
Today's GenAI class was about building a Voice Agent from scratch with @piyushgarg_dev sir.
The interesting part wasn't just making it talk it was understanding the architecture behind it.
Here's the flow 🧵👇
@Hiteshdotcom @nirudhuuu @ChaiCodeHQ
3
35
580
Streaming isn't just about getting tokens onto the screen.
The goal is a runtime where provider-specific complexity stays isolated while applications get a predictable streaming API.
Agni Agent SDK v1.2.0 🔥
Building the runtime one primitive at a time.
3
15
The complete pipeline:
Provider SDK
→ Provider Adapter
→ ModelStreamChunk
→ Runner
→ RunStreamEvent
→ Application
Each layer has a clear responsibility.
1
2
15
From the application side, streaming stays simple:
for await (const event of runner.stream(...))
Developers don't need Gemini's streaming objects or provider-specific parsing.
Just consume Agni's normalized events.
1
2
32
The Runner consumes the provider stream with for await...of.
Each ModelStreamChunk is mapped into a public RunStreamEvent.
The important boundary:
Provider-specific data stops at the adapter.
1
2
20
Providers speak different streaming formats.
Agni normalizes them into:
text_delta → generated text
tool_call_delta → tool-call data
done → stream finished
The Runtime works with one contract.
1
2
13
Streaming starts at the provider boundary.
ModelProvider exposes a common stream() contract.
Each provider can handle its own native streaming internally without forcing the Runtime to understand Gemini, OpenAI, Anthropic, or any provider-specific format.
1
2
28
Added streaming support to Agni Agent SDK
Agents no longer have to wait for the complete model response. Agni can now expose model output incrementally through a clean, provider-agnostic streaming API.
Here’s what I built ↓
@Hiteshdotcom @piyushgarg_dev @nirudhuuu
2
20
344
Today's GenAI class wasn't really about learning Inngest.
We first understood the problem Inngest is solving for developers building AI agents.
Queues, retries, background jobs, failures, orchestration...
Once the problem was clear, Inngest made a lot more sense.
🧵👇
2
29
622
Today's GenAI class marked an important milestone.
We completed our NotebookLM clone from scratch.
From authentication and workspaces to RAG, multiple data sources, and background processing we built the complete application.
But now comes the exciting part. 🧵👇
1
29
574
I'm also building the docs alongside the runtime.
Because an SDK isn't finished when the code works.
It's finished when another developer can understand it.
Agni is getting closer. 🔥
@Hiteshdotcom @piyushgarg_dev @nirudhuuu
@ChaiCodeHQ
#BuildInPublic #AI #TypeScript
5
83
Provider failures.
Tool failures.
Validation failures.
Guardrail blocks.
Configuration errors.
Each has stable error codes, metadata, and retryability.
The goal: make failures predictable and easier to debug.
1
5
80
Most AI agent demos end when the agent gives an answer.
I'm more interested in what happens when things go wrong.
While building Agni, I just finished the first structured error system for the runtime.
🧵👇
@Hiteshdotcom @piyushgarg_dev @nirudhuuu
2
21
745
The landing page is starting to look like the product I had in my head.
But the hard part isn't making an AI SDK look good.
It's making developers trust the runtime underneath it.
Predictable execution. Visible internals. Developer control.
The UI is just the first layer. 🔥
3
28
497
Still shaping the first impression of Agni Agent SDK.
Refining the landing page hero not just making it look good, but making sure it reflects the philosophy behind the SDK.
Building reliable AI agents, one step at a time. 🔥
@Hiteshdotcom
#TypeScript #AIAgents #BuildInPublic
3
1
44
1,075
<•> Agni
In Sanskrit, Agni means fire the force that transforms and carries intent. That idea inspired the name of what I'm building: Agni Agent SDK.
Not revealing yet... but if you're into AI agents, stay tuned.
The fire has just been lit. 👀🔥
@Hiteshdotcom @nirudhuuu
12
182
Today's GenAI class was all about getting hands-on.
We started building a NotebookLM-inspired application from scratch.
Instead of jumping into AI features, we focused on building a strong foundation first.
@Hiteshdotcom @surajtwt_ @piyushgarg_dev @nirudhuuu
1
21
337
Today's GenAI class was an introduction to Agent SDKs.
Instead of just learning how to use one, we built a simple Agent SDK from scratch.
Building it ourselves made the abstractions much easier to understand.
@Hiteshdotcom @piyushgarg_dev @nirudhuuu
2
33
493
Today's GenAI class was all about Memory in AI Agents.
One of the most confusing topics became much easier when we compared it to how human memory works.
Turns out, AI remembers a lot like we do.
Here's what I learned 🧵👇
@Hiteshdotcom @piyushgarg_dev @nirudhuuu
1
21
239
Today's GenAI class was about Vectorless RAG, PageIndex, and LLM Wiki.
Until now, I thought every RAG system needed embeddings and a vector database.
Turns out, that's not always true.
Here's what I learned 🧵👇
@Hiteshdotcom @piyushgarg_dev @nirudhuuu
1
23
305
Today's GenAI class was about Advanced RAG pipelines.
Learned new concepts and techniques beyond the basic RAG flow.
RAG looks simple on the surface, but production-ready systems require understanding many details.
@Hiteshdotcom @piyushgarg_dev @nirudhuuu
23
282
Built with Next.js, Gemini AI, PostgreSQL + pgvector, and the Twilio WhatsApp API.
This is v1 of a much bigger vision: a personal AI assistant with real long-term memory not just a chatbot that forgets everything after the session ends.
2
3
119
The feature I'm most excited about:
🧠 Memory resurfacing.
Mention something related to a past conversation, and DailyBrief connects it to what you shared before using semantic search + AI reasoning.
It doesn't just store information. It understands context.
1
2
45
Need a quick overview? Just send:
"brief me"
DailyBrief replies with everything that matters grouped, numbered, ready to act on. Reply "1 done" and it's handled. No app to open, no dashboard to check.
1
3
35
No new app. No folders. No manual sorting.
Just chat like you normally do on WhatsApp — DailyBrief automatically classifies everything into:
Tasks
Ideas
Notes
Reminders
Nothing gets lost in the scroll.
1
2
46
You don't have a forgetting problem.
You have a remembering problem.
Introducing DailyBrief — an AI memory assistant that works through WhatsApp.
Send messages naturally. It remembers what matters and brings it back when you need it.
3
1
22
773
You don't have a forgetting problem.
You have a remembering problem.
Your best ideas, tasks, and important thoughts are already somewhere in your WhatsApp chats.
But most of them disappear under hundreds of new messages.
3
1
23
734
Today we completed the streaming ChatGPT clone project.
The interesting part?
Calling an LLM API was only a small piece.
The real challenge was building everything around it:
Authentication, database, streaming, state management, and user experience.
@Hiteshdotcom @nirudhuuu
2
33
455
Today we started building a ChatGPT clone from scratch.
No AI responses yet.
Just laying the foundation that every production-ready app needs.
Here's what we set up and why it matters 🧵👇
@Hiteshdotcom @surajtwt_ @nirudhuuu @piyushgarg_dev
2
1
38
563
Today's GenAI class was about one of the most important concepts in building AI applications:
RAG
How do we make LLMs answer using our own data?
The answer: Don't retrain the model.
Give it the right information at the right time.
@Hiteshdotcom @piyushgarg_dev
@nirudhuuu
1
1
29
401
There's a version of AI that will teach you anything you want to know.
Including how to hurt someone.
No warning. No hesitation. No "I can't help with that."
Just an answer.
You've never met this version. Here's why 🧵
1
15
148
The most important concept in LLMs isn't the model.
It's the Context.
Too much context → hallucinations.
Too little context → poor output.
Today I learned SDKs, Agent SDKs, and the guardrails every AI engineer should know
@Hiteshdotcom @piyushgarg_dev @nirudhuuu
2
32
335
What's next:
conversation memory
streaming responses
RAG
more personas
building this in public, will keep posting progress. video attached above shows it in action 🎥
3
44
Biggest chunk of actual engineering time went into prompt engineering + token optimization.
Instead of one giant static prompt every request, it's built for dynamic few-shot example selection + modular prompt composition so token usage stays low but the persona stays consistent.
1
3
57
Architecture is fully modular:
persona management
prompt engineering
provider config
validation
chat generation
all separate modules. adding a new persona = plug in a new spec, not rewrite the whole thing.
1
2
47
Instead I actually researched the persona:
YouTube videos
live streams
X posts
blogs
personal website
Then converted all of that into a structured persona spec identity, communication style, teaching philosophy, vocabulary, response patterns, constraints, example conversations.
1
2
65
It's called Persona AI.
It recreates the teaching style & communication patterns of popular coding educators using LLMs.
Right now it can talk like Hitesh sir. Piyush sir is next architecture is already built to support multiple personas.
2
4
89
Built an AI that talks EXACTLY like @Hiteshdotcom sir 😭
not publicly available rn because it's a mythos-level persona that makes you a genuinely better engineer just by chatting with it
for more context check the first comment 👇
@piyushgarg_dev @nirudhuuu @ChaiCodeHQ
3
30
920
After months of curiosity…
Finally got the answer to why @piyushgarg_dev sir likes pink. 🌸😂
One mystery solved.
Back to shipping code. 🚀
28
456
GIGO: Garbage In, Garbage Out.
An AI's output is only as good as your input.
In LLMs, that input is called a prompt.
Learning AI isn't just about models it's about asking better questions.
Thanks @Hiteshdotcom & @piyushgarg_dev & @nirudhuuu
2
20
216
Everyone says learning AI means learning heavy math first.
Today, @piyushgarg_dev sir explained it with one simple analogy
ML Engineers build the engine. AI Application Engineers build the car.
Suddenly, Transformers didn't feel so scary.
Thanks @Hiteshdotcom & @nirudhuuu
2
15
228