Designing a Local-First Resume Parser: Architecture Decisions for Edge AI

The Problem I Want to Solve

Recruiters pay $0.05-$0.20 per document to cloud parsing services like DaXtra, RChilli, and Affinda. Beyond cost, there’s a privacy issue: candidate data leaves the device and travels to vendor servers.

I want to explore whether browser-based AI can eliminate both problems.

Project Status: Planning Complete, Build Starting

I’ve completed the architecture planning phase:

• PRD defining MVP scope
• 6 Architecture Decision Records (ADRs)
• 89-task implementation plan across 8 weeks

No production code yet. This post documents the planned architecture, not a working system.

Planned Technical Stack

Key Architecture Decisions

ADR-001: Leptos Over React

Decision: Use Leptos with SSR + hydration for the extension UI.

Tradeoffs:

• Pro: Single Rust codebase (UI + inference bridge + storage)
• Pro: Direct WASM memory access, no serialization overhead
• Con: ~500KB bundle vs ~100KB React
• Con: Smaller ecosystem, fewer examples

I’m betting that unified language benefits outweigh ecosystem maturity for this use case.

ADR-002: transformers.js + Phi-3-mini

Decision: Use transformers.js with Phi-3-mini q4 (~1.5GB quantized).

Why Phi-3:

• Optimized for instruction following
• 1.3B parameters fits browser cache
• WebGPU support in transformers.js

Risk: 1.5GB initial download. First-time UX will require patience.

ADR-006: WebGPU Primary, WASM Fallback

Detection logic:

async function selectBackend(): Promise<'webgpu' | 'wasm'> {
  if (navigator.gpu) {
    const adapter = await navigator.gpu.requestAdapter();
    if (adapter) return 'webgpu';
  }
  return 'wasm';
}

Expected performance:

• WebGPU: Target < 5s parse latency
• WASM SIMD: Target < 30s parse latency

These are targets, not measurements. Actual performance unknown until implementation.

MVP Scope

What I’m planning to build first:

• PDF drag-drop upload
• Client-side text extraction
• Local LLM inference with confidence scores
• Human-in-the-loop field correction
• Encrypted local storage
• JSON export

Not in MVP: LinkedIn overlays, ATS integration, team features.

Open Questions

Next Steps

Week 1: Extension scaffold with Leptos popup rendering. I’ll document what actually works vs. what was planned.

Documenting the build at Intent Solutions. jeremy@intentsolutions.io