Dipankar Sarkar
Founder, Neul Labs | AI Agent Security & Systems
Edinburgh, United Kingdom
Actions
Dipankar Sarkar is the founder and principal consultant at Neul Labs, a fractional AI CTO, and a hands-on applied AI engineer with 18+ years taking ambitious systems from research and architecture into production. His recent work focuses on governed tool-using agents, durable orchestration, evaluation, local inference, and AI-generated GPU-kernel correctness. Earlier, he architected regulated banking AI at Aveni, built RobotGPT at Orangewood Labs, and worked on NLP, recommendation, computer vision, and trust-and-safety systems serving more than 100M users at Hike. He is the author of two technical books, has published applied ML research, and holds an M.S. in Computer Science (Cybersecurity) from Arizona State University and a B.Tech. from IIT Delhi.
Links
Area of Expertise
Topics
Your AI-Generated GPU Kernel Passed the Notebook. It Is Still Wrong.
Attendees will leave able to turn a persuasive notebook demo into a reproducible correctness gate for AI-generated CUDA and Triton kernels.
An AI-generated kernel can compile, run quickly, and match a reference on a few friendly tensors while still failing on boundary shapes, strides, aliasing, dtype extremes, special values, or reduction order. This talk builds the verification workflow as a Jupyter notebook: an independent CPU oracle, operator-aware input generation, explicit tolerances, deterministic seeds, environment capture, failure minimisation, and an exportable evidence bundle.
The notebook compares two public controlled studies. In one measured corpus, the harness detected all 10 seeded LLM-style defects while all 16 correct controls stayed clean, with matching verdicts across five GPU classes. A separate 26-operator study shows why boundary-focused inputs and aggressive adversarial values trade recall against false positives.
The point is not that notebooks prove production safety. It is that a notebook can make the experiment inspectable, replayable, and portable enough for another person to challenge before apparent speed becomes trusted code.
Authorize the Effect, Not the Tool Name: Runtime Security for AI Agents
The outcome is an AI agent that can perform useful work without inheriting blanket authority, quietly repeating an ambiguous side effect, or leaving investigators with only a chat transcript. A tool called `send_invoice` or `rotate_key` is not a security boundary. The real decision includes the actor, purpose, target, parameters, materiality, time window, approval, and recovery path.
This technical session builds that boundary from first principles. We begin with a threat model for tool-using agents: prompt injection, confused-deputy actions, delegated sub-agents, retries after timeouts, identity drift, and escalation from one permitted action to a dangerous batch. A live demonstration then converts each proposed tool call into a canonical effect envelope, assigns an action tier, enforces short-lived authority and segregation of duties, makes irreversible effects idempotent, preserves a hash-chained decision receipt, and stops for a human when policy cannot decide safely.
The patterns draw on hands-on deployment of agents with Azure AI Foundry, custom sandboxing, and MCP in .NET, together with the public CloseGate and Regulus reference systems. Those systems demonstrate policy, approval, audit, and recovery patterns; they are not presented as certifications or universal security guarantees. The demonstration uses a mock high-impact action and no production credentials.
Intermediate security practitioners, architects, IAM and GRC professionals, and application-security engineers will leave able to:
1. Model an agent permission as an exact effect rather than a broad tool name.
2. Place human approval where identity, purpose, materiality, or irreversibility demands it.
3. Preserve an evidence record that distinguishes proposal, authorisation, execution, retry, and recovery.
4. Test prompt injection, delegation, replay, and batch-escalation paths before granting an agent real authority.
A Workflow Is Not an Agent: Where Determinism Must Win
Outcome: attendees can decide which work belongs in a deterministic automation, which genuinely benefits from an AI agent, and how to make the handoff safe.
This intermediate 40-minute live virtual architecture walkthrough follows one business request through intake, classification, proposed action, human approval, execution, receipt, and recovery. We use a Power Automate-style flow as the deterministic backbone and introduce an agent only where ambiguity earns its cost. Then we add the controls that demos omit: a typed handoff contract, action envelopes binding purpose and scope, least-privilege tools, idempotency keys, bounded retries, durable decision and outcome receipts, and escalation after an ambiguous timeout.
The session combines 18+ years of production software and distributed-systems work, current Azure AI Foundry agent and MCP in .NET delivery, and the workflow-design practice behind my 2026 Packt book, AI for Everyday Automation: 7 AI Workflows to Save Hours at Work Every Week. The patterns are vendor-neutral and illustrated with public reference systems, not presented as customer deployments. Attendees leave with a decision matrix, a handoff contract, and four recovery tests they can apply to Power Automate, Copilot Studio, or custom Foundry solutions.
Dipankar Sarkar
Founder, Neul Labs | AI Agent Security & Systems
Edinburgh, United Kingdom
Links
Actions
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top