Verdict
What it is
Verdict is an AI resume reviewer built with a privacy-conscious architecture. Resume parsing and embedding happen entirely in the browser using pdfjs-dist, mammoth, and a WASM-based embedding model (Xenova/all-MiniLM-L6-v2), so your full resume never leaves the device. Only the top-k semantically relevant chunks are sent to the LLM backend for scoring. It evaluates resumes on ATS compliance and quality (0–100), across 8 sub-score categories with actionable before/after suggestions.
The hard part
The hardest design decision was the privacy architecture: how do you run meaningful AI analysis while keeping the full resume off the server? Semantic chunking + RAG retrieval was the answer, but wiring up WASM embeddings in the browser with acceptable cold-load performance took significant iteration.
What it taught me
Building full-stack AI products taught me that the "AI" part is rarely the hard part: the data pipeline around it is. Client-side ML, IndexedDB storage, chunking strategy, and prompt engineering all had to work together seamlessly. Rate limiting and JSON parsing robustness are non-negotiable once you hit production.