founding software engineer / bangalore

himadri s chatterjee

senior software engineer / ai platforms / cloud systems

I build production AI platforms, real-time streaming services, payment infrastructure, and product-facing tools. Currently building things at Convai.

email resume github linkedin

experience

senior engineering range, startup pace

Mar 2022 - Present

Founding Software Engineer / Convai

Designed REST APIs and gRPC streaming services for real-time conversational AI. Owned early GCP infrastructure, autoscaling, stability and performance enhancements, Stripe subscriptions, and enterprise on-premise deployments.

  • Built core services powering real-time voice and text interactions at scale.
  • Designed the payment architecture for usage-based billing and secure checkout flows supporting millions in transaction value.
  • Successfully deployed GPU-enabled environments for enterprise accounts.
  • Co-built React creator-facing tools and integrations used across the product.
  • Helped with data processing and training for the in-house NeuroSync model powering character facial expressions.
May 2021 - Mar 2022

Systems Engineer / Tata Consultancy Services

Developed Azure infrastructure solutions and internal automation tools for enterprise operations, with a focus on compute provisioning, networking, identity, reporting, and resource tracking.

Jan 2021 - Mar 2021

Software Developer Intern / BNP Paribas

Built full-stack features for an internal invoice management platform using Angular, C#, and SQL Server.

selected projects

systems, ai, and product work

Compact notes on the work that best reflects how I think: system design, technical depth, and useful products.

ai infrastructure

Durable AI Agent Workflow Platform

A production-style platform that runs AI tasks as durable workflows. Hand-built execution engine with commit-before-execute steps, retry/backoff, idempotency keys for effectively-once side effects, and a crash-recovery sweeper. Adds human-in-the-loop approval gates, OpenTelemetry tracing to Jaeger, a cost/latency dashboard, and an automated eval harness across a benchmark suite.

FastAPICeleryPostgresRedisOpenTelemetryNext.js
github
ai systems / language models

GPT-2 from Scratch with LoRA + Speculative Decoding

A 124M GPT-2 built from scratch in PyTorch and taken end-to-end. Pretrained on 10B tokens of FineWeb-Edu across 4x A100 on Modal (val loss 3.30), then validated against OpenAI's GPT-2 on HellaSwag - 0.295 vs 0.311 acc_norm on the same harness. Adds LoRA fine-tuning with downstream benchmarks (SST-2 79.7%, E2E-NLG BLEU 21.6), a KV cache, and speculative decoding proven distributionally exact against plain sampling.

PythonPyTorchModalDDPLoRATransformers
github
agentic ai

Work Buddy - Workstream Control Agent

A multi-agent system (Google ADK) that turns messy project context - meeting notes, Slack, Jira, repo state - into an auditable execution ledger. Specialist sub-agents extract tasks, decisions, and blockers, reconcile duplicates across sources, and draft Jira/Slack/PR/email updates, with a human-in-the-loop approval gate before any external write, secret redaction throughout, and the same tools exposed over an MCP server. Built for the Google & Kaggle AI Agents Intensive capstone.

Google ADKMulti-AgentMCPGeminiFastAPI
github
video classification / 2019

Violent Crowd Behaviour Detection

Trained a C3D model on the ViolentFlow dataset to classify violent versus non-violent behaviour in crowd footage. The work was written up and published as a paper.

C3DVideoViolentFlowPublished
github
computer vision / 2019

Galaxy Zoo Morphology Classification

Final-year bachelor's project comparing convolutional architectures on the Galaxy Zoo dataset to pick the most suitable model for galaxy morphology classification. Published as a paper.

CNNAstronomyClassificationPublished
github

personal lab

side projects with a product lens

GTJ World

A personal product space at the intersection of web, community, and lightweight publishing. The goal is to keep the surface simple while making the underlying workflows useful and scalable.

ReactViteTailwindNetlify
gtj.world

Nightingale App

A live pet-care app that connects pet owners with trusted caretakers when they are away, and matches volunteers ("Nightingales") with pets that need company.

FlutterFlowFirebase
App Link

stack

from model internals to production infra

AI / ML

PyTorch, CUDA kernels, transformer architectures, multi-agent systems, LLM tool use, model data processing.

Cloud / Infra

GCP Cloud Run, GKE, IAP, Compute Engine, AWS, Kubernetes, Docker.

Backend / Data

Python, Go, gRPC, REST APIs, PostgreSQL, MongoDB, Firebase, Stripe payments.

Product Frontend

React, Next.js, JavaScript, HTML/CSS, creator-facing tools, integrations, and UI systems.

writing

notes, experiments, and technical essays

I keep longer-form technical writing on Medium, including older computer vision and machine learning notes. This section can later surface selected essays directly.

read on medium

contact

let's build systems people can actually depend on

I am open to senior software engineering conversations around AI platforms, cloud-native products, real-time systems, and product-minded engineering teams.

officialhimadri11@gmail.com github linkedin