Aryaman Todkar

I’m a software developer at Deutsche Bank, currently exploring the infrastructure behind real-time voice agents. I’m exploring and building around open-source voice AI and the systems problems involved in making conversational agents faster, with a focus on reducing end-to-end latency.

Outside software, I spend a lot of my time working out. I’ve been lifting for six years, do calisthenics, and run marathons. I also love traveling and have spent time across much of Southeast Asia. A longer solo trip is next on my list.

I’ve also always been drawn to startups. I like the idea of working in small teams where people move quickly, take ownership, and have a lot of agency. That’s the kind of environment I’d eventually like to experience.

Writing

All writing

Why My Side Project Failed to Become a SaaS OneGlanse worked as an open-source AI visibility tracker, but the browser and proxy infrastructure made a hosted SaaS too costly for one developer to operate. Oct 3, 2026 - 6 min read Why You Keep Putting Off What Matters How small unresolved possibilities compete for your attention, and what changed when I removed the easy ways to check them. Sep 26, 2026 - 3 min read How I Built an Open-Source ITN Model That Beat NVIDIA Thutmose on Voice-Agent Transcripts On a held-out synthetic benchmark, Premove reached 99.5% semantic accuracy by combining deterministic candidate generation with contextual neural scoring and exact decoding. Sep 12, 2026 - 11 min read

Projects

Premove ITN Open source context-aware inverse text normalization library - 2026

Premove ITN is an open-source, context-aware inverse text normalization system for conversational voice-agent transcripts, with open model weights. It turns ambiguous spoken forms into structured values such as numbers, times, dates, money, phone numbers, email addresses, URLs, versions, and identifiers.

It combines deterministic Python and Rust candidate generation, contextual DeBERTa scoring, and exact dynamic-programming decoding. It reached 99.50% semantic accuracy on the frozen voice-agent benchmark and 89.70% overall, with 56.49 ms mean warm latency on an Apple M4 MacBook Air.

GitHub PyPI Model weights

Premove ITN - resolve ambiguous speech before it reaches your tools

OneGlanse Open-source GEO tracker - 2026

OneGlanse is an open-source GEO platform that measures brand visibility inside real LLM interfaces rather than API outputs. It captures live ChatGPT, Claude, Gemini, Perplexity, and Google AI Overview responses, then turns them into time-series GEO reports.

Website GitHub Product Hunt