Writings
Thoughts on systems architecture, AI engineering, team building, and delivering outcomes that matter.
2026
Designing an On-Call Rotation That Works
Most on-call rotations burn out the engineers who are best at incident response, because the rotation punishes competence. Here is how to design one that doesn't.
Observability for AI Agents in Production
Standard APM tells you a request was slow. It doesn't tell you why an agent picked one tool over another. Here is what to log, trace, and alert on instead.
Trust Boundaries for Multi-Agent Systems
Single-agent guardrails like the Substrate Pattern don't compose automatically once agents delegate to other agents. Here is where multi-agent trust actually breaks.
Rust for Python AI Infra
Why Rust drop-ins accelerate Python-based AI infrastructure like LiteLLM proxies and LangGraph runtimes, and where the rewrite is and isn't worth it.
AI Agent Safety Governance Playbook
The practical playbook for AI agent safety governance. Threat model, runtime policy, prompt injection defense, audit, red-team, and the tooling matrix.
AI Architect in Financial Services
What I learned as Principal AI Architect at a UK fintech. Building production agent systems for financial services, the frameworks, and the lessons.
AI Governance: The 4-Tier NIST AI RMF
How to implement the 4-tier NIST AI RMF Agentic Profile for financial services. The framework I developed at Neul Labs for compliance-first AI agents.
Harmony Protocol in Rust: Parsing GPT-OSS
A deep dive into implementing OpenAI's Harmony response format in Rust. Channel semantics, streaming, comparison with ChatML, and the Rust implementation.
Rust for AI Agent Infrastructure
Why Rust dominates AI agent infrastructure. Memory safety, async runtime, the ecosystem, and the career path. A field guide from the founder of Neul Labs.
Building Developer Tools in the Open
Dipankar Sarkar shares lessons from 20 years of open source work, from Google Summer of Code to production developer tools that got adopted and sustained.
How Nykaa Scaled to IPO: Platform Migration
The engineering decisions behind Nykaa from startup to IPO — Magento to custom Python, in-memory cart service, Kong API gateway, 500% traffic growth, 99.99% uptime.
Measuring Engineering Impact That Matters
Dipankar Sarkar presents a framework for measuring engineering impact that connects technical work to business outcomes, beyond vanity metrics like story points.
LLM Provider Risk Management for Banks
How to manage LLM provider risk in regulated industries: the Trust Boundaries Model, multi-provider strategy, fallback patterns, and compliance mapping.
2025
Remote-First Engineering Teams That Deliver
Dipankar Sarkar explains how to build remote-first engineering teams that outperform co-located ones: communication architecture, async workflows, and trust.
Defence in Depth for AI Agents
A 5-layer kill switch architecture for production AI agents: application logic, tool execution, action policy, circuit breaker, and a manual switch.
AI Agent Governance for Financial Services
Dipankar Sarkar explains AI agent governance for banks: the Tiered Governance Model and Three Lines of Defence framework for SS1/23 and Consumer Duty compliance.
Hiring Engineers Who Ship: A Practical Guide
Dipankar Sarkar shares the hiring framework used to build teams at Nykaa, Hike, and Orangewood Labs: identifying engineers who ship, not just interview well.
Scaling AI-Assisted Development
Dipankar Sarkar explains how to scale AI-assisted development from small teams to enterprise. The Pivot Point framework for knowing when informal practices break.
AI Agent Safety: The Substrate Pattern
Dipankar Sarkar explains the Substrate Pattern for AI agent safety: building execution envelopes that constrain LLM-powered systems while preserving capability.
Invariants in AI-Generated Code
Dipankar Sarkar explains why unwritten invariants are the primary source of AI-generated defects, and how to make implicit constraints explicit in production code.
Production-Ready AI-Assisted Development
Dipankar Sarkar explains how to use AI code generation in production systems: the Vibes Inside Guardrails framework for AI productivity with reliability.
2024
Optimizing ARM Builds in CI/CD Pipelines
Practical strategies for optimizing ARM builds in CI/CD pipelines: emulation challenges, native build setup, cross-compilation, and cloud-native approaches.
EdgeML and the Future of Robotics
Dipankar Sarkar built a next-generation robotics SDK at Orangewood Labs using EdgeML for on-device AI processing, enabling intelligent robotic systems.
AutoInspect and AutoSpray: ML for Robotics
Dipankar Sarkar built AutoInspect and AutoSpray at Orangewood Labs: ML and computer vision systems for industrial quality control and precision spray painting.
Machine Learning at Hike: Five AI Systems
Dipankar Sarkar led development of five ML systems at Hike serving millions: computer vision avatars, vernacular NLP, social matchmaking, and trust & safety.
Real-Time Data Ingestion for E-Commerce
Dipankar Sarkar built Nykaa's real-time data ingestion framework processing 5B+ daily events using AWS Lambda and Apache Flink, enabling data-driven personalization.
2023
RoboGPT: Natural Language Robot Programming
Dipankar Sarkar developed RoboGPT at Orangewood Labs, an LLM-powered interface for natural language robot programming in industrial automation.
A Real-Time Personalized Feed for E-Commerce
Dipankar Sarkar built a TikTok-inspired real-time personalized feed at Nykaa using AI/ML, achieving 200% increase in daily active users and 150% more engagement.
An Integrated Ad Platform and Social Commerce
Dipankar Sarkar built Nykaa's integrated ad platform and social commerce solution, achieving 200% revenue increase and onboarding 1000+ influencers.
2022
2021
Scalable E-Commerce Infrastructure at Nykaa
Dipankar Sarkar led Nykaa's platform migration from Magento to custom Python, implementing in-memory cart service and Kong API gateway for 500% traffic growth.
Analytical Systems in Core Java at Tyroo
Dipankar Sarkar built analytical systems at Tyroo (India's largest AdTech company) processing 10B+ daily ad impressions using Core Java for real-time insights.
2020
2012
Early Days at SlideShare, Before LinkedIn
Dipankar Sarkar joined SlideShare as one of its first five engineers, building core features for the platform later called the "YouTube of presentations."
PHIRE: Mobile Banking at MPower Money
Dipankar Sarkar built PHIRE at MPower Money - the world's first mobile debit network enabling SMS-based banking for unbanked populations using Python.
2011
2010
Vision Algorithm Research at B-Core Software
Dipankar Sarkar optimized computer vision algorithms at B-Core Software Tokyo, gaining international experience in image processing and algorithm optimization.
A Web Service PDS Plugin for Oracle Reports
Dipankar Sarkar developed a Web Service PDS plugin for Oracle Reports during internship at Oracle Corporation, receiving recognition from US headquarters.