Rahul Sunil

I work on the unglamorous half of applied AI.

Most of what makes an LLM system usable in an enterprise has very little to do with prompt tuning. It is whether the document you uploaded was actually parseable. Whether the citation in the output points at real text or at something plausible the model invented. Whether tenant A's evidence can ever surface in tenant B's assessment. Whether a run that takes twelve hours can be made to take one. Whether you can prove, before you cut over, that the new code path produces the same answers as the old one. That is the layer I architect.

At Certa, my primary responsibility was architecting the complete production system for Risk Agent — taking it from early concepts into a reliable, enterprise-grade compliance product. Rather than just focusing on model accuracy or prompt logic, I owned the overall system architecture: the agent orchestration, production infrastructure, service design, scalability, multi-tenant vector isolation, and operational reliability required to process vendor evidence at scale. Today, Risk Agent evaluates vendor SOC 2, PCI DSS, and pentest reports across 21 risk domains and 840 controls, replacing 10–12 hours of manual analyst review per assessment with under 10 minutes of automated execution. My contribution wasn't just making an AI capability accurate — it was architecting the complete production system that made that capability usable as a real enterprise product.

Working on Risk Agent, Contracts Agent, and helping architect Data Agent revealed a clear pattern: every new agent team was rebuilding the same foundational plumbing from scratch — separate ingestion pipelines, separate tenant isolation schemes, and separate operational infrastructure. Olympus was not created because someone explicitly requested it; it emerged from those production lessons as a platform-level engineering initiative. Designed to power more than six agents once live, Olympus provides production-grade capabilities — unified agent execution, tenant isolation, error taxonomies, and cost/usage telemetry — from Day One. Instead of spending months building infrastructure, Product and Engineering teams can take new agent ideas into production in weeks, focusing their effort on business logic and agent-specific behavior while building on a flexible, reusable foundation.

I applied that same end-to-end production ownership to Contracts Agent, taking it from an n8n proof-of-concept idea into a fully production-ready service in approximately 1.5 months. I owned the complete path to production: designing the standalone microservice architecture, implementing async job models with idempotency guards and bounded retries, building structural multi-tenant isolation in Qdrant, and establishing operational infrastructure for load balancing, monitoring, and uptime. The service was deployed directly into our first enterprise client's live contracting workflow, while a subsequent pipeline optimization cut LLM inference cost per contract by roughly 60%.

Earlier at Certa, my work on DesignAI marked a major turning point, helping pivot the company toward an AI-first direction that contributed to securing $35M Series B funding. Facing a fundamental architectural challenge in workflow automation, I tackled a monolithic, non-deterministic AI routing approach and transformed it into a deterministic logical routing system with well-defined classification. By taking an opaque prompt apart into three structured classification problems behind a routing layer, accuracy leaped from 62% to 97%. This work led to a patent filed in the US and India as sole inventor and earned a co-win of Certa's first-ever company-wide Genius Award, demonstrating my experience identifying fundamental AI architecture bottlenecks and making complex AI systems deterministic, structured, and controllable.

Before Certa, I was a research intern and project lead at Microsoft, leading a team engineering non-invasive wildlife-monitoring systems using unmanned ground vehicles — authoring long-range telemetry libraries and field robotics software. I also co-founded AgriVision, building early computer vision and data pipelines for crop-disease detection, and developed multi-tenant security microservices at Accubits.

I spoke at Google DevFest 2024 about running large language models in production, contributed to the PyCon India 2023 website, and serve as a member of the Kerala Police Cyberdome technology group.