H2LooP is building AI coding tools for the software hidden inside hardware
The Bengaluru startup makes hardware-aware AI models for firmware and embedded systems, and has raised $2 million to take them into chips, cars and defence.
Every drone, car and chip runs on software most people never see: the low-level firmware and drivers that sit closest to the hardware. That code is slow and costly to write, still leans on manual reviews and hand-written compliance documents, and can swallow months of a senior engineer’s time each time a product moves to a new chip.
General AI coding assistants, tuned for web and application software, are little help here, because they do not understand a processor's register map or its timing constraints. H2LooP, a Bengaluru startup, is building the AI that does.
The company was founded in May 2025 by Sairanjan Mishra and Pulkit Agrawal. Both had started a company before. Mishra, the CEO, spent more than 15 years building software for complex systems at Philips, Toshiba, Cisco and Bosch, and knows how slowly these industries buy. Agrawal, the CTO, spent a decade at Google building low-level software.
Their shared view, in Mishra's words, is that in these systems, writing code is rarely the hard part. Understanding it, working out how it touches the hardware, and proving it is correct is where teams lose months.
Small models, and a map of the hardware
H2LooP's answer pairs two things. The first is a set of small language models trained specifically on embedded-systems code, rather than the general internet text most models learn from. The second is a proprietary knowledge graph that ties hardware specifications, safety standards, design patterns and a customer's own codebase into a single map the models can reason over.
The point of the pairing is context. An AI that writes a device driver without understanding the target chip will produce code that looks right but fails on integration. A verification layer then checks that generated code is correct and auditable before it reaches production.
The whole system can run on-premises or fully air-gapped, so a customer's code never leaves its control, which matters a great deal in defence and semiconductors.
Live in chips and cars, aiming at drones
H2LooP sells to regulated, mission-critical industries: semiconductors, automotive, aerospace, telecom and defence, and increasingly data centres and robotics. It is already deployed with semiconductor and automotive teams across India and Europe.
At one German semiconductor company handling AUTOSAR compliance and legacy-code conversion, the company says its tools cut design-update and audit time and generated test cases with 90 to 95 percent accuracy. Across deployments, it reports roughly a doubling of engineering velocity. It has also been selected for the Infineon Global Startup Program and recognised by the India Electronics and Semiconductor Association.
In early 2026, H2LooP raised $2 million in a seed round co-led by Speciale Invest and 3one4 Capital, to strengthen the platform, scale enterprise deployments, and push into higher-complexity areas like data centres, drones and robotics.
The money arrived at a good time. Investment in Indian AI rose sharply through 2025 and into 2026, and companies building hardware for cars, chips and defence all face the same software bottleneck H2LooP works on. The founders make a national case for it: India has long written the world's most critical software, and they want it to build the AI layer that makes that software faster and safer to trust. Each new deployment, they say, helps the next, because every customer's code and test results make the models sharper.
(This story has been researched and compiled using publicly available information.)