Self Improving Agent Framework & 1M Context LLMs - Workshop on Recursive Agents and Deep dive into DeepSeek-V4 Architecture
Welcome to a technical workshop laying out the architecture of DeepSeek-V4 and teaching you how to deploy a 1M reasoning agent as well as how to optimize LLM Training.
Speakers
This might be the most detailed and up-to-date workshop you can find. Our speakers have taught about half a million engineers around the globe.
Hamza Farooq (Stanford & UCLA Faculty on AI, Founder @ Traversaal)
Zain Hasan (Together AI, 200,000 learners on Udemy, Coursera)
Ali Shafique (AI Research Scientist @ Traversaal.ai)
Session 1— Self Improving and Recursive Agents(12 pm)
Presented by Hamza Farooq and Ali Shafique.
Design a skill-based agent architecture: Structure a root orchestrator, a skill library, and specialist sub-agents that run in parallel with no hand-coded routing rules.
Understand recursive execution: See how a complex customer complaint gets decomposed, delegated concurrently, and synthesized into a coherent resolution.
Train without a dataset: Learn how resolution outcomes (CSAT, ticket closed, escalation avoided) become dozens of training trajectories automatically through the execution tree.
Apply the RAO reward design: Understand why every node in the tree gets its own reward signal and why that's what makes agents learn to delegate well, not just delegate more.
Know the build path: Phase 1 runs today on Claude API , no GPU. Phase 2 swaps in an open-source model and trains on real outcomes with GRPO.
Bottom Line: Attendees leave knowing how to build agents that decompose hard problems, coordinate specialists, and get measurably better over time, with a clear path from prototype to production.
Session 2 — Inside DeepSeek-V4: Architecting 1M Context for Agents and Reasoning (1.15 pm)
Presented by Zain Hasan.
Deconstruct Hybrid Attention — the mechanics behind DeepSeek-V4's efficiency
Compress the KV Cache 10x — strategies that slash long-context inference cost
Deploy 1M-Context Workloads — full-repo coding agents, multi-step reasoning, the things the old context window couldn't hold
Build Long-Horizon Agents — coherence across massive trajectories without falling apart
Schedule
11:30 — Doors & coffee
12:00 — Recursive and Self Improving Agents (Hamza & Ali)
1:00 — Break & Lunch
1:15 DeepSeek Architecture Deep dive (Zain)
2:00 — Adjourn