Advitya Sood

I'm a senior full-stack engineer with 6+ years building products end to end — system design through backend, frontend, and production ownership.

I'm currently at Decisions, a Microsoft Teams–integrated meeting productivity platform, working remotely with a team in Oslo. Two things there represent my work best. I architected and shipped a centralised licensing and entitlements service from the ground up, covering multi-tenant access control, role-based authorisation, and subscription state, which replaced manual tenant provisioning. And I built the generative AI suite — AI Recap, Workflow Assistant, AI Minutes — orchestrating LLM workflows on Azure OpenAI across streaming and non-streaming inference paths, with document ingestion pipelines that parse, chunk, and segment meeting attachments. It's now the most-used feature in the product. The stack is Angular with NgRx and Signals on the frontend, C# / ASP.NET Core and Azure Cosmos DB on the backend, with Microsoft Graph for calendar, file, and identity operations.

Before Decisions I led frontend engineering for an EDR cybersecurity platform at EclecticIQ in Amsterdam, where the problem was keeping large real-time and historical security-event views responsive under sustained streaming load — profiling, virtualised rendering, and change-detection tuning. Before that I was the founding technical hire at Coingale, a crypto trading platform, building live charts over WebSocket feeds across multiple exchanges. Earlier I led application delivery across UN agencies — the Office of Counter-Terrorism, UNICEF, and OICT — on systems used by governments and serving millions of daily active users, and I co-founded a project presented at the UN General Assembly in New York.

What I'm most interested in now is production AI systems — specifically the engineering rather than the modelling. Non-deterministic output on a write path is a genuinely hard problem: a model that hits its token ceiling returns a truncated response that looks structurally valid, and a naive implementation persists it. Getting that right means treating model output as untrusted input, with explicit validation and idempotency guards before anything writes. I like that class of work, and I like owning things after they ship — most of my production debugging is distributed tracing and KQL against Application Insights.

I'm looking for senior full-stack or AI-platform engineering roles, ideally where I own a surface area end to end rather than a single layer. Based in New Delhi and comfortable working remotely across European and US-overlapping hours.