Thiyagarajan Maruthavanan, known as Rajan, is co-founder and chief executive of Kalmantic Agentic Lab, which develops infrastructure for running AI agents with greater control over inference costs, reliability, and deployment. A former Intuit product leader and co-founder of the Indian SaaS accelerator Upekkha, he argues that successful AI businesses must eventually own the infrastructure supporting their products.
From computer vision to AI infrastructure
Maruthavanan studied computer science at the International Institute of Information Technology, Hyderabad, and founded a computer-vision startup in 2007. At Intuit, where he held product leadership roles from 2011 to 2017, he worked on QuickBooks Online and the challenges of adapting accounting software across markets.
He subsequently co-founded Upekkha, helping Indian software founders develop capital-efficient SaaS businesses, find product-market fit, and expand internationally. He later became its chief executive as the accelerator shifted toward AI-powered software. His writing for founders emphasizes solving specific customer problems instead of mistaking access to powerful models for a defensible business.
While building Ultasono, an application that infers which text prompt might have generated a song, Maruthavanan confronted rising inference expenses firsthand. He moved the application from Anthropic-hosted models onto NVIDIA DGX Spark hardware, then began running agents locally. Memory limitations and enterprise reliability requirements exposed the practical difficulties of replacing rented inference with owned infrastructure.
At Kalmantic, which he founded with Kashi KS and Ananya George, he has developed jusInfer infrastructure for agents requiring persistent memory, scheduled execution, and human oversight. His approach to inference ownership permits rented model access while startups validate demand, but favors infrastructure control after product-market fit, when rate limits, vendor dependencies, audit requirements, reproducibility, and recurring token costs become material.
- LegacyCodeBench: Maruthavanan co-authored this COBOL understanding benchmark, which evaluates AI-generated documentation through claim-based behavioral verification: extracting claims about program behavior and testing them against execution of the original code.
- PeakWeights: Co-authored with Vamshi Ambati, this open-source quantization project identifies important model parameters without requiring a calibration dataset, addressing the constraints of running capable models on limited hardware.
Maruthavanan also describes himself as the author of Peak Inference: Infraeconomics of AI Inference, extending his focus on the financial and operational implications of deploying AI systems.
