Manifesto
A model that works in a notebook is a different problem from one that has to answer a phone call.
I did not come to AI systems from research. I came from shipping, and from finding out repeatedly that the model is the small part. The rest is latency budgets, state that survives a pod restart, tenants that must never see each other's data, and a handoff to a human that cannot drop the conversation.
That is the work I do: the whole system around the model, designed, built and deployed.
Production voice bots built on a streaming speech-to-text, LLM and text-to-speech pipeline, then migrated to Gemini Live's native audio models. Call recordings are generated for every conversation.
Caller → STT → LLM → TTS → Caller
Chatbots that pass a live conversation to a human agent. Redis pub/sub carries the transfer between Kubernetes pods, so it does not matter which pod holds the customer's connection and which holds the agent's.
Bot → Redis pub/sub → Agent, any pod
Database and dashboard design for onboarding voice-agent customers: tenant isolation, role-based access control and the analytics dashboards each customer sees.
Tenant → Roles → Agents → Analytics
Docker and Kubernetes deployment behind an API management gateway, with API authentication, IP restrictions, a Redis cache and blob storage for recordings and assets.
Gateway → Auth → Services → Cache / Storage
Garment try-on and face makeup try-on, built as computer vision pipelines at a generative AI startup.
Photo → Vision model → Composite
Generative pipelines that turn a prompt or a single image into a 3D asset.
Prompt or image → Generation → 3D asset
Practice
The model is one component. I build the rest of the system around it.
Speech-to-text, LLM and text-to-speech agents, native-audio models, call recording and live handoff to people.
Try-on, image and 3D generation, and the pipelines that move pixels between models.
Chatbots, retrieval over private data, and adaptive products where an LLM drives the experience.
Kubernetes, API gateways, caching, storage, authentication, access control and multi-tenant data design.
Own product
An AI learning advisor I designed, built and run myself. A voice guide called Iris assesses where you are, then the product builds a learning path around the answer.
Architecture notes are on the way.
Long-form breakdowns of decisions that held up in production. Leave an email and I will send them as they are published.
About
Full-stack, then machine learning, then 3D and vision, now realtime voice in production.
I started as a freelance full-stack developer straight out of a BTech. A year in I moved into data science, first classical machine learning and then deep learning, building models for clients who cared whether it worked on their data.
In 2023 I joined a startup working on 3D generative AI and computer vision. Since 2025 I have been building production voice AI: agent architecture, Kubernetes deployment behind an API gateway, Redis-backed state, multi-tenant data and live human handoff.
Contract work, remote roles and technical advisory. Based in India at UTC+5:30, which overlaps European mornings and US evenings.