Tathagat

AI Engineer · Himalayan region of India

I build intelligent systems and like understanding exactly how, and why, they work.

Selected work

• Production computer-vision inference on Triton

  Vigyanam Labs · product under NDA · 2025 — present

  A facial-analysis screening product needed to run several vision models and a multimodal LLM on every submission, with throughput and reliability good enough for both a consumer app and a multi-tenant clinical B2B product. The product itself is under NDA, so this study describes the engineering, not the product.

• Fine-tuned LLMs for a proprietary language, served with vLLM

  Pronto Software, Australia · via Netsmartz · 2024 — 2025

  Pronto's product is written in its own proprietary programming language. Developers wanted to ask questions about the codebase in plain English and generate code in that language, without any source code leaving their infrastructure.

• Backend architecture and asynchronous workflows for AI products

  Netsmartz and Vigyanam Labs · 2024 — present

  AI features tend to arrive as notebooks. Turning them into dependable services means auth, rate limits, async I/O, background work, feedback capture, structured logs and deployment, repeated across several products.

• Acoustic gearbox fault detection

  Visiting engineer, via Netsmartz · 2024 — 2025

  Tell whether a running tractor gearbox is healthy, humming, rumbling or chattering from its sound and vibration, so faults are caught before assembly is complete.

• Should-be-costing from CAD drawings

  Visiting engineer, via Netsmartz · 2024 — 2025

  Given an AutoCAD drawing of a tractor gear, estimate what it should cost to manufacture, so procurement can negotiate from evidence.

Essays

• Why exported models disagree with the notebook (2026-10-17)
• Gear-mesh frequency, in one line (2026-10-17)

Now (2026-10-17)

Building: This observatory, and the production hardening of a GPU Triton inference service at work. Studying: CUDA and distributed training: DDP, FSDP, NCCL, and profiling training loops.

Experience

• Vigyanam Labs — Technical Director, AI · Lead AI Engineer (Dec 2025 — present)
• Netsmartz — AI/ML Engineer (Jul 2024 — Dec 2025)
• Salt Holdings — Data Analyst Intern (Jan 2024 — Jul 2024)

Things I’m exploring

• CUDA: To understand what a kernel actually costs, instead of trusting a framework's number.
• Distributed training: The models I serve are increasingly trained across many GPUs; I want to reason about that, not just configure it.
• DDP and FSDP: Where the memory goes, and what sharding really trades away.
• NCCL: All-reduce is the heartbeat of multi-GPU training. I want to see it, not assume it.
• Profiling and performance: I already profile inference services in production; now I'm extending that to training loops.
• Mathematical foundations: Linear algebra, probability and optimisation, revisited slowly with pen and paper so the intuition is mine.

9 photographs in “A moment of calm”.

Contact

tathagat1@outlook.in · LinkedIn