CraftifAI

Crafting the physical world with Agentic AI platform

[CraftifAI @craftifai]

Watch us build live object detection on a TurtleBot and motor control firmware from a datasheet, right in front of you.

We're at TechEx Events Amsterdam, today & tomorrow booth T27.

PipeGen runs the whole vision pipeline on the TurtleBot itself. Nothing goes to the cloud.

FirmGen takes a datasheet, writes the motor control firmware, then builds and flashes it onto the board in front of you.

We build each one live and walk through what it's doing, so you can see how it works.

Worth stopping by if you're an embedded or firmware engineer who has spent too many weeks on board bring up, a computer vision engineer trying to get a model off a workstation and onto a device, or a robotics engineer working on motion control and perception at the same time.

Leading an engineering or hardware team at an OEM or industrial firm? It's a quick way to see what this looks like on real hardware.

Bring a board or a datasheet and try it on your hardware.

Know someone at TechEx? Send this their way.

#TechExEurope #TurtleBot #Firmware

[Google for Developers India @GoogleDevsIN]

@GoogleStartups @Adalat_AI @Aikenist1 @Aurassure @craftifai @buildonnodeops @dodopayments @h2loopai @OnFinance_ai @PipeshiftAI @potpiedotai @proxgy @soundverse_ai @SuperBryn_AI @TartanHQ @securezeron Ayna, Binocs, Fitsol, FlexifyMe, Jidoka Technologies

[CraftifAI @craftifai]

Every silicon company ships a reference design and a BSP. But almost none know how long a customer actually takes to ship a product.

Looking back at SEMICON India this week, design in time was all people cared about.

When eval kits go out, there is a long, quiet stretch. Teams are forced to write firmware by hand against new SDKs and unread datasheets.

The cost of that stretch lands entirely on Field Application Engineers (FAEs).

They find out why customers stall. It’s almost never the silicon, it's a peripheral configured wrong, or a power sequence buried in an app note. FAEs solve it one customer at a time. The next customer hits the same wall, and the cycle repeats.

That massive, quiet stretch is where FirmGen sits.
It captures the hardware context first: pin
assignments, addresses, power constraints then generates firmware against that specific silicon revision.

The industry theme this year is "silicon to systems."
The middle of that phrase is still done by hand.

#SemiconIndia2026 #embeddedsystems #firmware

[CraftifAI @craftifai]

Day 2 at #ElectronicaIndia2026 #ProductronicaIndia2026, and we realized most teams face the same hurdle.

Nobody is starting from a blank repo. They all have a 4-year-old codebase, a shipping product, and a roadmap demanding they add new sensors to it.

Which led to the same 3 questions about our AI agents:
1. "Will it rewrite everything?"
No. It reads the existing structure and makes targeted changes.

2. "Does it support our other MCU families?"
Hardware context is declared once. The abstraction sits across families.

3. "Does our code leave the building?"
On-prem, private cloud, or cloud. For enterprise, that passes the security review immediately.

The key idea here is that AI for firmware isn't about writing code from scratch. It’s about capturing your pin assignments and power constraints first, so the output actually matches your board.

The day also brought a nice moment: Messe Muenchen India recognized CraftifAI as part of their India's Powerplay in Electronics campaign.

One more day at Hall 2, Stand H2.F29. Come find us.

#electronics #EdgeAI #firmware

[CraftifAI @craftifai]

Day 1 at #ElectronicaIndia2026 #ProductronicaIndia2026, and people were already flabbergasted by the leaps in hardware automation.

We brought the bench to the floor.
PipeGen drove the TurtleBot demo, inference running on the module itself with nothing leaving the device.
FirmGen handled the other one: datasheet in, working firmware out for a machine vibration node, on a board visitors could stand over and watch boot.

Having people ranging from Industrial IoT teams to robotics builders and silicon engineers all give the exact same reaction tells us one thing.

The massive gap we’ve been working to solve between hardware that exists and software that runs on it is finally ready to close.

If you are at the show, come find us at Hall 2, Stand H2.F29 and see it boot for yourself.

#embeddedsystems #robotics

[Pratik Sharda @pratiksharda2]

Digital twins are table stakes now. That's the good news.

The gap that still eats weeks: the twin's depth/collision world is clean, and the real camera on an Orin is noisier, slower, and thermally constrained twenty minutes into a run.

Most teams discover this after the policy looks great in sim.

We're spending a lot of PipeGen time closing exactly that sim-to-device perception gap — same pipeline graph, compiled for the hardware you actually ship.

[Pratik Sharda @pratiksharda2]

Most Physical AI stacks are fighting one clock rate.

Reasoning can sit at ~10 Hz. The camera→model→control loop cannot. Obstacle and depth perception need to stay on-device, inside a few tens of ms, or the slow model ends up in the safety path.

That's the split we design for with PipeGen at CraftifAI: agents compile the hardware-aware perception pipeline for the Jetson. Cloud/VLA stays advisory. The fast path stays deterministic.

If it doesn't run inside the latency budget on the board, it doesn't count.

[Pratik Sharda @pratiksharda2]

Something about embedded we haven't talked about yet.

We ran the entire buildathon on agents — from the first application to the final score.

[Pratik Sharda @pratiksharda2]

Ran our first EMBEDDED buildathon in Bangalore last Saturday,

on what turned out to be the busiest event day of the year here with multiple events and hackathons running at the same time, and 100+ engineers still showed up.

600+ registrations came in over the two weeks, solid turnout for a first run.

Gave them 4 hours. Solo. Real boards, our agents, mentors on the floor.

Engineers built and flashed on the boards in 4 hours.

Embedded doesn't get hackathons because setup alone eats a weekend. That's the part we wanted to break.

First Orbit. More coming.

[Pratik Sharda @pratiksharda2]

Folks found a safe way to mining, no more accidents due to lack of communication

@craftifai #EmbeddedHackathon
@ST_World

[Pratik Sharda @pratiksharda2]

100+ people in the house for @craftifai EMBEDDED Hackathon at @MicrosoftIndia Bengaluru!

Time to go all out with the most innovative ideas on bare metal…let’s goo 🚀

[Google for Startups @GoogleStartups]

Meet the 2026 class of the Google for Startups Accelerator: India!

20 AI-first startups. ~2,500 applications. A milestone cohort as Google's Accelerator programs celebrate 10 years 🚀

The cohort reflects the deepening technical maturity of the ecosystem, solving high-impact challenges across climate, healthcare, developer workflows, and finance. Over the next three months, these startups will refine their products, optimize their architecture, and prepare for scaled deployment.

🥳 Congratulations to @adalat_ai, @Aikenist1, @aurassure, Ayna, @binocs_labs, @craftifai, @dodopayments, @FlexifyMe, @Fitsol_green, @h2loopai, @jidoka_tech, @NodeOpsHQ, OnFinanceAI, @PipeshiftAI, @potpiedotai, @proxgy, @soundverse_ai, SuperBryn, @TartanHQ, and @securezeron.

Learn more about the cohort here: https://t.co/Qtc76RKbpK

[Varun R @RoboticswithV]

Embedded Hackathon @Microsoft Bengaluru!

Register Here : https://t.co/JWmCCLCHtw

@devaarambh @craftifai

[Pratik Sharda @pratiksharda2]

Buildathon is next Saturday (8th Aug). Quick note on the format since people keep asking.

Pick a track. Get a board. Build all day. Demo live at 6pm.

No slideware round. If it doesn’t run on the hardware, it doesn’t count.

Evaluation by the team that built the agents. Winners get cash, everyone who ships gets swag.

Small room on purpose. We’d rather have 100 people who build than 500 who network.

RSVP: https://t.co/sF8c6MdRog

[Pratik Sharda @pratiksharda2]

Hosting a hardware + AI buildathon in Bangalore on Aug 8

We hand you a jetson or an esp32. our agents handle the firmware and the perception pipeline. You just build.

The bet: give devs agents for the plumbing, see what ships in 9 hours.

50k prize pool. Hardware provided. @Microsoft office.

RSVP here - https://t.co/hxfeGdf8aa

[CraftifAI @craftifai]

Hiring: Physical AI System Architect at CraftifAI (Bengaluru, Hybrid)

If you've deployed perception pipelines on Jetson, or custom edge hardware in production, not just in a notebook, we want to talk.

5-10 yrs. Edge AI. Robotics. Real-world deployment.

Apply: https://t.co/im4bx0ASJ6

#hiring #PhysicalAI #edgeAI

[CraftifAI @craftifai]

CraftifAI has been selected for the @GoogleStartups Accelerator among the top 20 teams out of 2,500+ applicants.

Building the infrastructure layer for Physical AI is a hard problem. Firmware generation, edge perception pipelines, silicon design automation. The companies that win the next decade of robotics and autonomous systems won't just have great hardware. They'll have the software stack that makes it move.

That's what we're building with CraftifAI. Access to @Google 's engineering depth, AI research, and global network at this stage is a serious accelerant.

On to what's next.

[Pratik Sharda @pratiksharda2]

We got featured at @GoogleIndia’s office in Bangalore 🚀

Looks like a mini times square billboard moment

Thank you @GoogleStartups!

[3one4 Capital @3one4Capital]

Last Wednesday, 3one4 Capital and AWS brought engineers, researchers, founders, and investors together in Bengaluru for a builder session on physical AI. The Model to Machine: A Builder Session on Physical AI event centered on the challenge of scaling advanced research into commercial applications. (1/4)

[CraftifAI @craftifai]

You stack detection, tracking, and classification on one Jetson. Each model works perfectly in isolation.

Run all three concurrently and GPU memory allocation becomes a silent war. The tracker steals memory from the classifier's batch buffer during peak load. The classifier output arrives 2 frames late. The detection model drops to 22 FPS because the memory scheduler is context-switching between three inference streams that were never profiled together.

Nobody benchmarks multi-model memory behavior. Everyone benchmarks models one at a time and assumes they'll coexist.
They don't.

#EdgeAI #DeepStream