Agents, sandboxing, multi-agent collaboration, automated evaluation and security. The gap between something that works on your laptop and something you would let a stranger use.

Five build days, two live sessions, one ship day, one read day. Monday, October 12 to Sunday, October 18.

What this week is actually for

Weeks 1 and 2 built a system that answers. Week 3 builds one that acts, then puts every guard around it that acting requires.

The order matters. You build the ReAct loop on Day 15, then on Day 16 you immediately sandbox it, because an agent that can execute code on your machine without isolation is not a feature, it is an incident. Days 18 and 19 are the two things most AI portfolios are missing entirely: evaluation and guardrails.

If you are aiming at enterprise or client work, this is the week that qualifies you.

Day Date Focus Type
15 Mon, Oct 12 ‣ Build
16 Tue, Oct 13 ‣ Build
17 Wed, Oct 14 ‣ Build plus Live Session 5
18 Thu, Oct 15 ‣ Build
19 Fri, Oct 16 ‣ Build plus Live Session 6
20 Sat, Oct 17 ‣ Ship Day
21 Sun, Oct 18 ‣ Read Day

Live sessions this week

Session 5, Wednesday October 14, 7:30 PM IST, 90 minutes. Two agents that argue until the work is good enough, plus the cost math on screen.

Session 6, Friday October 16, 7:30 PM IST, 90 minutes. We inject a hidden instruction into a document, watch the agent obey it, then build the guardrail that stops it.

Tech Support runs every day at 9:00 PM IST on Discord.

What you have at the end of the week

A ReAct agent with a hard step ceiling, an isolated execution sandbox, a Coder and Reviewer pair with an approval gate, an automated eval suite scored by an LLM judge, and an input and output guardrail layer. Plus a 60-second recorded demo.

That is Portfolio Build 3 of four.

Note on the sandbox

Day 16 builds subprocess isolation with memory limits and a timeout, not WebAssembly. WASM and Docker isolation are covered as concepts and appear in the interview questions, because interviewers ask about them, but the build you ship is subprocess-based.

That is deliberate. Subprocess isolation is what you can write and test in ninety minutes. Understanding why WASM is stronger is a separate thing from shipping it.