Building a Browser-Based Semiconductor Wafer "Autopsy Lab" with Kiro (Kiroween 2026)
Written as part of my Kiroween 2026 hackathon submission: Chip Autopsy Lab. Live Demo: https://chip-autopsy-lab.netlify.app GitHub Repo: https://github.com/suyin31/chip-autopsy-lab Demo Video: https://youtu.be/h5vQuV
Written as part of my Kiroween 2026 hackathon submission: Chip Autopsy Lab.
- Live Demo: https://chip-autopsy-lab.netlify.app
- GitHub Repo: https://github.com/suyin31/chip-autopsy-lab
- Demo Video: https://youtu.be/h5vQuVODXHA
Why I Built This
I'm an IC design undergraduate currently training as an ATE test engineer. Whenever a wafer lot comes back with bad yield, small teams usually end up writing one-off Python scripts, exporting to Excel, and plotting one wafer at a time in matplotlib β taking hours at 2 a.m. just to answer a single question: why is it failing?
Commercial yield analysis tools cost tens of thousands of dollars, while free browser viewers only draw wafer maps without diagnosing the root cause. Worse, uploading raw STDF binary test logs to third-party servers isn't an option under fab/OSAT NDAs. For Kiroween 2026, I wanted to build a local-first, offline-capable tool that opens raw STDF v4 files right in the browser and leads with a plain-sentence verdict backed by clickable die-level evidence.
How Kiro Changed My Development Workflow
On a project mixing binary parsing, Web Workers, Canvas2D rendering, and a defect classifier, it's easy for core logic and UI state to tangle together. Building with Kiro in a spec-driven workflow kept the architecture clean from day one:
1. Specs + 14 Correctness Properties First
Before writing feature code, I set up .kiro/specs/chip-autopsy-lab/ (requirements.md, design.md, and tasks.md) with 14 explicit correctness properties, each paired with fast-check property-based tests (src/core/**/*.prop.test.ts).
STDF v4 is unforgiving β optional trailing fields, inherited PTR limits, multi-site interleaving, and truncated streams from Teradyne/Advantest testers. Having round-trip property tests derived directly from Kiro specs caught subtle parser bugs early, long before they could show up on a wafer map.
2. Steering Rules to Lock Architectural Boundaries
Using .kiro/steering/ (product.md, tech.md, structure.md, design-language.md), I enforced strict boundaries across every agent turn:
-
src/core/stays framework-free TypeScript (ESLint blocks React, Next.js, and DOM imports) and runs inside a Web Worker; - Raw STDF bytes never leave the user's browser β even the optional AI narrator only receives aggregate facts.
3. Hooks as an Always-On Safety Net
I wired three save hooks in .kiro/hooks/:
-
lint-typecheck-on-save.kiro.hookβ runs ESLint + typecheck on everysrc/**/*.ts(x)save; -
core-unit-tests-on-save.kiro.hookβ runs colocated unit and property tests onsrc/core/**/*.tssaves, with a rule never to weaken a design property just to pass; -
i18n-key-parity-on-save.kiro.hookβ verifies translation key parity acrossen,zh-CN, andzh-TW.
Results
-
Local-first STDF v4 engine: parses 16 record types (plain or
.gz) in a Web Worker with truncation recovery; - Verdict-first UI: clicking any "cause of death" finding highlights the exact dies on the Canvas2D wafer map and its bar in the bin Pareto;
- WM-811K Spatial Defect Classifier: 97.8% accuracy (macro-F1 0.885) on 51,886 held-out real wafer maps;
- Verified quality gates: 474 unit and property tests passing, plus Playwright e2e and frame-rate performance gates.
This was my first hackathon, and Kiro's spec-driven structure + automated hooks made it possible to ship an industrial-grade tool solo. Try out the live demo or check out the repo on GitHub!
Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.