AI, LLM agents, MCP

Likely to come up as a conversation ("how do you use AI?") rather than a quiz. Have a clear, honest position.

How you use AI in daily work

I use GitHub Copilot and Roo Code every day, and Claude Code on my own projects. I treat the agent like a fast junior pair: I give it context and constraints, it drafts, and I review every line. Tests are the judge, not the agent's confidence. It's great for boilerplate, exploring an unfamiliar library, writing tests and first drafts of docs. Design decisions, review and anything touching security or money stay with me. I've built reusable skills, saved instructions with our conventions and checklists, so the agent's output matches team standards.

Can you trust AI-generated code? How do you review it?

Not blindly. Review it like any PR, more strictly where it's plausible-but-wrong: off-by-one errors, missing edge cases, invented APIs, outdated library usage, security (injection, secrets), and concurrency. Require tests, and run them. Keep changes small so review is real.

What are reusable skills / custom instructions?

Saved instruction files the agent loads for a type of task: coding conventions, how to run tests, a review checklist, a migration recipe. They make outputs consistent across a team and save re-explaining context. Prayog's CLAUDE.md is an example: engine invariants (integers for money, one writer thread, no system clock), test-first, "measure, don't guess", and a session workflow that ends with explain-back questions.

Does AI make engineers less necessary?

It shifts the job toward specifying, reviewing, designing and verifying. Knowing why a system works matters more, because you have to catch the agent when it's confidently wrong. That's why I learn domains deeply; my CV calls it "a curious generalist who learns fast, which matters more in the age of AI".

Concepts

What is an LLM agent?

An LLM in a loop with tools: it reasons about the goal, calls a tool (API, database, search), observes the result, and continues until done. Components: the model, instructions, tools with clear schemas, memory/context, and guardrails. Multi-agent systems split responsibilities across specialist agents with an orchestrator.

What is MCP?

The Model Context Protocol, an open standard (introduced by Anthropic, now widely adopted) for connecting AI applications to tools and data. Servers expose tools, resources and prompts; clients (IDEs, agents) discover and call them over JSON-RPC (stdio or HTTP). One integration works with any MCP-capable client. In the hackathon, Fi Money's MCP server exposed net worth, credit report, EPF, mutual funds and bank transactions.

Prompt engineering vs. context engineering?

Prompt engineering: how you phrase instructions (role, task, constraints, examples, output format). Context engineering: deciding what information is in the context window for each step: relevant data only, tool results trimmed, history summarised, the right examples. Most agent failures are context failures: missing information or too much noise.

RAG vs. fine-tuning?

RAG retrieves relevant documents at query time and puts them in the prompt: good for private or changing knowledge, with citations, no training. Fine-tuning changes the model's weights: good for style, format or a narrow task, but expensive and the knowledge goes stale. Start with prompting, then RAG, then fine-tuning.

How do you reduce hallucinations?

Ground answers in retrieved or tool data, ask the model to cite sources, allow "I don't know", constrain output (schemas, enums), lower temperature, validate outputs in code, and evaluate with a test set.

How do you evaluate an LLM feature?

A fixed set of representative inputs with expected outputs or rubrics; automatic checks (format, facts from tool data), LLM-as-judge for open-ended answers, human review of samples; run on every prompt or model change, track cost and latency too.

What are the risks of agents that take actions?

Prompt injection (malicious text in data telling the agent to do something), over-broad permissions, irreversible actions, data leakage. Mitigate: least-privilege tools, human confirmation for risky actions, treat tool output as untrusted data, logs and audit trails.

Bedrock vs. Vertex AI vs. calling a model API directly?

Bedrock (AWS) and Vertex AI (Google Cloud) give many foundation models behind one managed API, inside your cloud account's security, networking and billing, plus agent and RAG building blocks (Bedrock Agents and Knowledge Bases; Vertex AI Agent Builder and Google ADK). Calling a provider's API directly is simpler and gets new features first. Enterprises often pick the cloud platform for governance.