The problem
Tailoring a portfolio project to a specific job posting is days of work: read the company, pick a project that proves the right skills, build it well, publish it credibly. LetsBuild automates the whole path — a job description goes in, a production-shaped GitHub repository comes out, with README, CI/CD, Architecture Decision Records, and a realistic commit history spread across days instead of one mega-commit.
Architecture
Ten orchestrated layers, each with a narrow contract: an agent harness (middleware chain, Docker sandbox, compiled policy gates), a JD intake engine with guaranteed structured output, company intelligence running six parallel research sub-agents, six-dimension weighted matching with gap analysis, a project architect that designs skill-driven projects with ADRs, a multi-agent Code Forge, a GitHub publisher, a content factory for blog and video material, memory, and deterministic agent hooks enforcing policy at every boundary.
Two design decisions carry the safety story. Every line of generated code executes inside a Docker sandbox before anything is published. And review is genuinely independent: the reviewer agent receives zero context from the coder agent — no shared conversation, no shared rationale — so it cannot rubber-stamp reasoning it never saw.
Learning is explicit rather than emergent. The ReasoningBank runs a RETRIEVE → JUDGE → DISTILL → CONSOLIDATE loop over past runs, distilling what worked into retrievable guidance for future ones. The repo also ships AgentForge Arena, a tournament mode where five-agent teams (architect, builder, frontend, tester, critic) build against the same challenge head-to-head, ranked by Bradley-Terry ratings on an ELO leaderboard.
What's measured
The architecture is fully documented — 23 named production patterns, a data-flow reference, and a 100-step execution plan — and the arena produces head-to-head outcomes that make agent-team configurations comparable rather than anecdotal.
Honest limitations
The README advertises ATS match scores and a self-learning improvement curve; those are design targets without a published benchmark behind them, so no such number is repeated here. The deeper question — whether generated portfolio projects actually influence hiring outcomes — is untested by construction. This page states what the pipeline does, not what it wins.