Saksham Dhull
Software Engineer (Team Lead), AI Agents, Wisdom AI
Bengaluru, India
Actions
I lead the AI Agents team at WisdomAI, where I build production AI agent systems and work on agent architecture, reliability, evaluation, and efficiency. Previously, I worked at Samsung Research HQ in South Korea on multimodal AI and 5G/6G systems, and later co-founded an AI startup. My work spans AI agents, LLM systems, distributed systems, and production ML infrastructure. I hold a B.Tech. in Computer Science from IIT Delhi.
Area of Expertise
Topics
Keeping Context Small: Systems Lessons for AI Agents
As agents grow, their context grows with them: more tools, instructions, memory, and state. Loading all of that on every request is expensive and often unnecessary.
This talk looks at how to keep agent context small using ideas from systems design: working sets, caching, lazy loading, and dynamic tool discovery. We’ll cover both the theory and the practical trade-offs, when cached reads are much cheaper than normal input, what should you cache, how much should you preload, and when should tools be fetched on demand? The goal is to reduce token cost and context bloat without hurting agent quality.
Engineering Predictable Systems with Non-Deterministic AI
LLMs are non-deterministic, but production AI agents still need to feel predictable. If two users ask similar questions, or an agent works through the same kind of problem twice, the exact wording may change, but the decisions, actions, and overall outcome should stay reasonably consistent.
This talk looks at how to build that kind of functional consistency around non-deterministic models. I’ll cover where variability enters an agentic system, how it compounds through reasoning and tool use, and some practical ways to keep behaviour within predictable bounds using constrained execution, state management, validation, retries, tool design, and evaluation.
Designing Tools for Agents
Most APIs and tools are designed for and by the developers. AI agents don’t use them the way humans do. Small choices around tool granularity, naming, schemas, descriptions, outputs, and error handling can have surprisingly effects on how well an agent reasons and acts.
This talk looks at patterns that work and patterns that don’t when designing tools for AI agents. We’ll cover when to split or combine capabilities, how much logic should live inside a tool, how to return useful information without polluting context, how to design for failures and retries, and how to evaluate a given tool. The goal is to build tools that make agents reason, work and progress better.
Agents: Engineering Non-Deterministic Software
An AI agent can take different paths through the same task. How do we build one that completes useful work reliably, safely and without unnecessary cost?
This session starts with the fundamentals: what an agent is, how models, instructions and tools work together, and when an agentic loop is worth using. We then walk through the design of a simple coding agent and gradually extend it from writing standalone scripts to exploring repositories, modifying code and investigating bugs.
Each extension introduces new engineering challenges. We’ll examine how to design useful tools, evaluate outcomes, manage growing context, plan longer tasks and recover from failures. We’ll also cover prompt injection, permissions and sandboxing, before exploring how model selection, execution modes and caching can reduce latency and cost.
The focus is on understanding design decisions rather than learning a particular framework. Attendees will leave with a practical approach to building a first agent, diagnosing its limitations and improving it through measured experiments.
Audience: Computer science students and developers interested in building AI agents. Basic programming knowledge is expected; prior experience building agents or training models is not required.
Format: A lecture-style walkthrough using one evolving coding-agent example, diagrams and brief pseudocode. This is not a hands-on workshop, and attendees do not need to install anything.
Preferred duration: 50 minutes, followed by 5 minutes of Q&A.
AWS Student Community Day Mysuru 2026 Sessionize Event Upcoming
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