Public reference systems + delivery experience 01
Governed agents
High-impact agent actions remain permissioned, reviewable, interruptible, and explainable.
Applied AI systems · St Andrews, Scotland
Make the AI system dependable, not just impressive.
I’m Dipankar Sarkar. I help teams move ambitious AI from research or prototype into systems people can operate: governed agents, durable workflows, exact evaluation, local inference, and cloud-to-edge execution.
● Founder, Neul Labs Fractional AI CTO Remote-first from Scotland
Current question
How much authority can an AI system hold while people remain meaningfully in control?
- years building production systems
- 18+
- user scale for ML systems at Hike
- 100M+
- seeded kernel defects caught in a measured corpus
- 10/10
- companies founded across 2008–2024
- 6
The work
Four outcomes worth engineering for
The technology changes quickly. The operational obligations do not: authority, correctness, recovery, and ownership must be designed into the system.
01
Let agents act without surrendering control
Put identity, purpose, permissions, approval, evidence, and a stop path around tool-using agents before they cross into real side effects.
Evidence: Regulated banking AI, CloseGate, and Regulus
02Make failure visible and recovery deliberate
Treat an agent as a distributed system: explicit state, durable checkpoints, idempotent effects, replay, evaluation, and operator-owned recovery.
Evidence: fast-langgraph and production platform delivery
03Measure correctness, not plausibility
Use exact and operator-aware oracles where generated code can compile, benchmark well, and still be wrong.
Evidence: gpuemu and two public 2026 preprints
04Move intelligence closer to data and action
Design the boundary between cloud planning, local inference, edge execution, physical constraints, telemetry, and human responsibility.
Evidence: Cognisoc, RobotGPT, and WareMax
Selected work
Systems and papers you can inspect
Runtime governance, durable agent state, and exact evaluation for generated GPU code—implemented, measured, and documented in public.
Open source + arXiv preprint 02
Durable and coordinated agent systems
Agent work survives crashes, concurrent writers, and the distance between a decision and its side effect.
Open source + arXiv preprints 03
Exact evaluation for generated code
A fast generated kernel is useful only after the team can establish that it is correct.
Books & playbooks
Teach the system clearly enough that someone else can operate it
Two Packt books span web infrastructure and everyday AI workflows. What Generative AI extends the teaching into open GenAI and Agentic AI playbooks for business and technical leaders.
Packt book 2026
AI for Everyday Automation
7 AI Workflows to Save Hours at Work Every Week
Connected workflows for email, meetings, research, reports, task tracking, spreadsheets, and professional communication.
Read more → Packt book 2011
Nginx 1 Web Server Implementation Cookbook
More than 100 recipes for Nginx infrastructure
Configuration, application serving, reverse proxies, load balancing, caching, monitoring, and security for Nginx.
Read more → Apache-2.0 playbooks 2026
The GenAI and Agentic AI Playbooks
21 open chapters in ten languages
A business-leader guide to GenAI adoption followed by an engineering and operating guide to agent systems.
Read more → Where this is useful
Technical depth, translated into an operating outcome.
I work across the line from research and architecture to code, evaluation, production controls, and the team that will own the system afterward.
→Financial services and regulated AI
→Robotics, warehousing, and manufacturing
→Cloud platforms and developer tooling
→Consumer machine learning at scale
→Blockchain infrastructure and protocol security
→Startups, venture building, and technical diligence
Talks and workshops
For people who have to make AI work on Monday morning.
I turn the same evidence into vendor-neutral talks, technical podcasts, panels, and hands-on labs for engineering and industrial audiences.
View speaking topicsAn AI Agent Walks into Production
State, side effects, recovery, and the stop button
The Agent That Couldn’t Wire Money
Enforcing least privilege, approvals, and replay across tools
Your GPU Kernel Passed the Vibe Check. It Is Still Wrong.
Testing generated CUDA and Triton with operator-aware oracles
A useful first conversation
Start with the outcome that has to change.
Tell me what people need to be able to do, what currently fails, and what evidence would justify the next investment. We can work backward from there.