Small is beautiful: H2Loop’s opportunity to spark an SLM-led AI silicon revolution

There’s a quiet but profound transformation building at the hardware-software boundary — below the applications layer where most AI productivity tools play, and deep in the firmware that actually makes silicon do things. This is where H2Loop is staking its claim.

H2Loop’s recent close of a $2 million seed round, co-led by Speciale Invest and 3one4 Capital, validates a thesis we’ve held for some time: the next frontier of productivity lies not in general-purpose models, but in domain-specific, AI-native infrastructure for the physical world.

For too long, the engineers writing safety-critical lower level embedded & firmware code for automotive ECUs and avionics have been left out of the AI wave. Application developers enjoy GitHub Copilot; systems engineers have been left with reams of datasheets and manual code reviews. In mission-critical domains, code that merely “looks” right can become a liability — generic LLMs hallucinate register names, confuse vendor-specific API conventions, and produce failures that are costly to catch and dangerous to miss.

As manufacturers race to deliver software-defined vehicles and autonomous aerospace platforms, the complexity of underlying firmware is ballooning. The inability of current AI tools to reason about hardware constraints has opened a multi-billion-dollar gap in the systems engineering market — one that threatens to slow the global hardware innovation cycle as software development falls behind rapid silicon advancement.

“Hardware innovation is advancing rapidly, but the software powering it needs to keep pace in the age of AI. H2Loop is addressing a deeply technical and underserved problem by building AI-native infrastructure for systems software — an area that will be foundational to the next wave of hardware breakthroughs. — Arjun Rao, Founding Partner, Speciale Invest

The strength of H2Loop’s enterprise rests on the provenance of its leadership. CEO Sairanjan Mishra brings fifteen years of technical experience directing production-grade releases at Philips, Cisco, Toshiba, and Bosch. Co-founder Pulkit Agrawal is an AI/ML expert, with a prior decade at Google refining low-level, performance-critical software systems. Together, they bring a combined half-century of expertise in the very engineering H2Loop is now automating.

Their technical approach pairs domain-specific Small Language Models (SLMs) trained exclusively on low-level systems code with a proprietary context engine  — creating a hardware-aware semantic layer that general-purpose models fundamentally lack. This allows the AI to reason about specific register configurations and hardware specifications directly from technical documentation. The results are already compelling: Their Spark Preview, a 7B-parameter model, beats frontier systems on 8 of 13 embedded code categories, with a 70.4% in-domain perplexity reduction. Smaller, faster, and purpose-built.

In sectors like defence and semiconductors, where intellectual property is the most valuable asset, deployment architecture matters as much as model performance. H2Loop is engineered for on-premises or air-gapped environments — proprietary code and hardware specs never leave the secure perimeter. This “AI sovereignty” creates a flywheel effect: every deployment enriches a proprietary context engine, deepening the platform’s understanding of engineering intent and building a moat that cloud-dependent generalist competitors cannot cross.

Early enterprise deployments have already demonstrated a 200% improvement in development velocity and up to 95% accuracy in automated test case generation for legacy code conversion.

We invested in H2Loop because they are not building another code generator — they are building code assurance. By bridging the gap between silicon specifications and software implementation, H2Loop is positioned to become foundational infrastructure for the era of Physical AI. A generational opportunity, built in India, scaling globally.