Mumbai Tech Week 2026: My Key Takeaways from Two Days at BKC

Masterclasses- The Session that I Attended

The Future of AI: From Research to Reality (Meta AI) 

This session dived into three key technologies:

• SAM3 — the next generation of Meta’s Segment Anything Model
• Omilingual — multilingual AI capabilities
• PyTorch — advances in the open-source deep learning framework

Beyond Context Engineering: The Case for Domain-Specific LLMs 

This was one of my personal highlights. The session challenged the popular assumption that bigger models are always better — the “Bigger Model Fallacy”.

Key topics:

• FiMI — Financial Model of India. building domain-specific models
• AI Model Assistance for Chat — practical approaches to deploying LLMs in real workflows
• The 4 Stages of SLM Fine-Tuning:
  1. Base model selection
  2. Continuous pre-training
  3. Supervised fine-tuning
  4. Alignment

The core argument: for many enterprise applications, a well-tuned small language model (SLM) outperforms a generic large model. Context engineering alone isn’t enough — domain specificity matters.

Build with Codex

A deep dive into OpenAI’s Codex framework:

CODEX = Model + Harness + Surfaces

Key features:

1. Plugin — a combination of Skills, MCP, and APP
2. Automation — task execution without manual intervention
3. Sub-Agents — nested agent workflows for complex tasks
4. Worktrees — parallel copies of your codebase, enabling multiple agents to work simultaneously on different branches

The session also covered the Codex Security Plugin — safeguards for keeping agentic coding within defined boundaries.

Agentic AI for Enterprises: Building Secure, Production-Ready Systems

This session framed the shift from prompt engineering to a full AI Operating System mindset:

1. Build your Claude OS
2. Replace prompts with files
3. Let Claude think for you
4. Add tools
5. Automate everything

It introduced the concept of AI-DLC (AI Development Lifecycle) — treating AI deployment with the same discipline as software engineering.

Hiring Engineers in the AI Era: The New Interview Loop (Day 2)

A provocative and practical session on how engineering hiring must change. Instead of testing syntax, evaluate:

1. AI fluency — how well a candidate works with AI tools
2. Judgement and taste — knowing when AI output is good or needs correction
3. Code review at speed — reviewing AI-generated code efficiently
4. Systems and cost — understanding infrastructure and cost implications of AI
5. Debugging and verification — catching what AI gets wrong

The Engineering Contract: What Every Engineering Team Needs from Agentic Systems (My favourite of the event)

This was the masterclass I’d recommend to every engineering leader. It introduced a powerful mental model: instead of vague instructions, define explicit contracts between your team and your AI agents.

Old Contract vs New Contract — the shift from informal prompting to structured, documented agreements.

The five contract types:

1. System Contract (Agent.md)

• Custom commands
• Explicit boundaries
• Things the agent must never touch
• Specific code style requirements

Key insight: “Contracts are not essays.” Longer files are token-expensive, counterintuitively worse in performance, and create bloat. Keep them tight.

2. Skill Contract Enables the AI to:

• Gain domain knowledge
• Access specific tools
• Understand success criteria
• Operate within defined constraints

3. Memory Contract Defines:

• What gets remembered
• Where it lives
• Who writes it
• How long it persists
• What gets forgotten

The context ceiling problem was also highlighted — long-running agents inevitably hit a context limit, and Memory Contracts help manage what survives that ceiling.

4. Review Contract — criteria for when and how agent output is reviewed

5. Escalation Contract — when the agent must stop and ask a human

The session closed with a look at multi-tasking and Worktrees (/multitask) and the emerging Marketplace and Plugins for Cursor.