Three tracks. All shipping.

What to build

• Agents that call real tools and APIs, not just chat
• Multi-step planning with memory and recovery from errors
• Human-in-the-loop checkpoints where a mistake would be costly

Example problem statements

1. A research agent that reads sources, compares them and writes a short, cited brief
2. A support agent that resolves tickets by calling real tools, with a human sign-off for risky actions
3. A personal ops agent that schedules, follows up and keeps track across email and calendar

Judging focus

• Task success on realistic scenarios
• Sound planning and tool use
• Safety: limits, confirmations and fallbacks

What to build

• Event-driven workflows across services
• Data pipelines that clean, enrich and route information
• Monitoring and alerts that catch problems early

Example problem statements

1. Turning messy emails and form responses into clean, routed tickets
2. Spotting duplicate or suspicious records before they reach a database
3. Log and metric monitoring that explains what broke and alerts the right person

Judging focus

• Measurable time or error saved
• Reliability: retries, idempotency, edge cases
• Clear observability

What to build

• Retrieval-augmented features grounded in real data
• Natural-language interfaces to search and actions
• Multilingual experiences for users across India

Example problem statements

1. Natural-language search over a large set of documents, like “which policies changed since March?”
2. A study or writing assistant that works in English and Hindi
3. A help centre that answers from docs and FAQs, with sources

Judging focus

• Grounded, accurate answers
• Evaluation: how you measured quality
• UX, latency and cost awareness