AI agent engineer. Built an agent communication SDK with ~1K GitHub stars. I ship production workflows across email, Slack, WhatsApp and voice. https://xitter.cf/t.co/nEZ9corDLw
BITS PILANI
Joined February 2025
- Tweets1.8K
- Following759
- Followers1.3K
- Likes4.4K
Pinned Tweet
Hello, world 👋
I build AI agents and the infrastructure around them.
> Built an open-source Agent Communication SDK with ~1K GitHub stars and 6K+ installs across npm/PyPI
> Google Summer of Code @ Kubeflow
> Exploring agentic systems + AI × hardware
Over the past few months, I’ve gone deep into agents: single agents, workflows, memory, tools, orchestration, multi-agent systems and swarms.
Along the way, I’ve built:
> Production agent systems for startups and enterprise clients, including internal agentic infra for a ~$150M company
> An AI user-analytics platform combining product usage, payments, errors, support and other user signals
> Real-time multimodal agents using voice, vision and tools
> Agent infrastructure across communication, integrations, memory, tool calling, CLI and MCP
> Projects at hackathons, including the Google DeepMind Bangalore Hackathon
Before agents, I worked across backend, cloud and infrastructure. In college, I helped build systems serving 5K+ users, then worked across Kubernetes, MLOps, payments, DevRel and startups.
Recently, I’ve started exploring AI × hardware as a side project by building a conversational desk robot / personal agent.
Mostly here for:
AI Agents • Agent Infrastructure • OSS • AI × Hardware • Things I’m building
More: madmecodes.in
inspiring! love this no-bs approach.
i think “fake it till you make it” has its place too, sometimes you just need that push to realise your potential. but faking numbers or straight-up lying is a no-no.
But yhh no-bs for the win!
when we raised our seed round years ago, i actually told investors exactly how we *could* fail.
i'm generally a no bs person, so i wanted to lay all the cards out. the pitch had 3 parts:
-here's what we think this could become
-here's what we know so far
-here are all the risks for how this company could fail
i told them in a meeting, "you could tell me 5 risks you see. i'll tell you 100, because we've done the work to understand all 100."
we raised $4.5m a couple of weeks later, and the journey started.
investors aren't stupid. they hear pitches all day, and they can tell when a founder is hiding the downside.
so if you're looking to raise: don't oversell the upside, and definitely don't hide the downside. honesty wins.
— written with @WisprFlow
Ayush Gupta retweeted
The speed and urgency with which you execute and get things done tells everyone how badly you want a certain outcome;
Yapping is useless unless it pulls in more money and attention.
Ayush Gupta retweeted
Hot take: being super intelligent at work and super stupid at love gives you the best life
I love people who don’t rely on Placement Units 😍😍
I hope this PU concept dies in India too. Apply directly, reach out, build your own network, and create your own opportunities, just like they do abroad.
Ayush Gupta retweeted
spent the last month building Limbo.
it woke me up, ordered my lunch, ran my computer, browser, and raised my PRs while I was in bed. watch it.
introducing Limbo - a physical AI companion for work and daily tasks.
we've spent the last month building something playful and personal. an agent you enjoy interacting with, that still gets the work done.
it remembers you, uses your computer, browser, connects to smart devices, makes phone calls, drafts emails, orders food, create PRs and much more.
have a look at it yourself, and you'll know 👇
spent the last month building Limbo.
it woke me up, ordered my lunch, ran my computer, browser, and raised my PRs while I was in bed. watch it.
introducing Limbo - a physical AI companion for work and daily tasks.
we've spent the last month building something playful and personal. an agent you enjoy interacting with, that still gets the work done.
it remembers you, uses your computer, browser, connects to smart devices, makes phone calls, drafts emails, orders food, create PRs and much more.
have a look at it yourself, and you'll know 👇
Ayush Gupta retweeted
introducing Limbo - a physical AI companion for work and daily tasks.
we've spent the last month building something playful and personal. an agent you enjoy interacting with, that still gets the work done.
it remembers you, uses your computer, browser, connects to smart devices, makes phone calls, drafts emails, orders food, create PRs and much more.
have a look at it yourself, and you'll know 👇
Ayush Gupta retweeted
One of our customers is building a new product and planning for 10,000 sandbox instances. So they did the hosting math.
Even at 6% CPU utilization which was their average usage, Vercel Sandboxes were 15× more expensive than Agent37.
And at 10,000 instances, that a $731,000 bill with vercel vs $47,600 with agent37.
If you're looking to scale agents and needs sandboxes, Agent37 is a no-brainer.
Ayush Gupta retweeted
OpenClaw and Hermes are proof the grey market scales.
@openclaw's WhatsApp "scan the QR" setup runs on Baileys, an unofficial, reverse-engineered version of WhatsApp's own protocol.
How @NousResearch's hermes calls you on WhatsApp:
1. A sidecar built on whatsapp-rust links to a spare WhatsApp number as a linked device, same as WhatsApp Web.
2. It speaks WhatsApp's own encrypted protocol, including the call layer: ringing, accept, encrypted audio.
3. Call audio is piped to Gemini Live for real-time speech.
4. When it needs to do real work, Gemini hands off to the Hermes agent and speaks the result back.
Okay, Hermes making phone calls is unreal.
Hermes doesn't have to wait for you to check in anymore. It can just call you.
A new plugin in the Hermes catalog, hermes-whatsapp-voice, gives Hermes live voice calls over regular WhatsApp. No business account needed.
What that looks like:
▸ It rings you when something actually needs you, like an urgent reminder or a scheduled morning check-in
▸ You can call it back and have it handle things mid-call
▸ When you hang up, a recap lands in your WhatsApp chat
▸ It can only ring your number. Nobody else's
Notifications pile up and get swiped away. A ringing phone gets answered.
It's community-built and runs on an unofficial WhatsApp connection, so set it up on a spare number.
OpenClaw and Hermes are proof the grey market scales.
@openclaw's WhatsApp "scan the QR" setup runs on Baileys, an unofficial, reverse-engineered version of WhatsApp's own protocol.
How @NousResearch's hermes calls you on WhatsApp:
1. A sidecar built on whatsapp-rust links to a spare WhatsApp number as a linked device, same as WhatsApp Web.
2. It speaks WhatsApp's own encrypted protocol, including the call layer: ringing, accept, encrypted audio.
3. Call audio is piped to Gemini Live for real-time speech.
4. When it needs to do real work, Gemini hands off to the Hermes agent and speaks the result back.
Okay, Hermes making phone calls is unreal.
Hermes doesn't have to wait for you to check in anymore. It can just call you.
A new plugin in the Hermes catalog, hermes-whatsapp-voice, gives Hermes live voice calls over regular WhatsApp. No business account needed.
What that looks like:
▸ It rings you when something actually needs you, like an urgent reminder or a scheduled morning check-in
▸ You can call it back and have it handle things mid-call
▸ When you hang up, a recap lands in your WhatsApp chat
▸ It can only ring your number. Nobody else's
Notifications pile up and get swiped away. A ringing phone gets answered.
It's community-built and runs on an unofficial WhatsApp connection, so set it up on a spare number.
Exploring @superset_sh
> I like its terminal first approach
> And the fact i can have multiple PRs worktree, agents working simultaneously.
I will leave Warp for coding tasks and use it just for go to terminal...
Built a pipeline using inverted indexes, fine-tuned embeddings, LightGBM, cross-encoders, and Qwen.
Here's how I built it, what worked, and what didn't.
Ayush Gupta retweeted
My company is hiring SWE interns!
A few basics we’re looking for:
• A clear understanding of basic system design
• Familiarity with coding agents and how to use them
• An understanding of plugins, connectors, and how they fit into your workflow
Interested? DM me your resume or portfolio.
I went through a 176-setting study on AI agent harnesses.
The finding? Trimming old context before summarizing it can save costs without hurting performance.
Here's what I learned and how I'd apply it to MCP agents.
Blog link: madmecodes.in/blog/context-m… View article
Independent auditability is the one people skip. if the agent can write its own logs, you don't have logs.
The 7 principles he lays out:
1. model diversity: no model checks its own work
2. observe everything: tamper-proof evidence of every action
3. verifiability: test failures and attacks, not just happy paths
4. independent controls: the org decides access, not the model
5. independent auditability: the model never controls its own evidence
6. containment: assume compromise, keep a kill switch
7. incident disclosure: report failures and share the fix
