Portfolio OS — Universal Data-Driven Personal Operating System

Problem

Engineering portfolios are typically static, fragile markdown sites or hardcoded templates that fail to demonstrate real systems architecture, cannot dynamically prove claims with grounded evidence, and require high backend hosting costs for AI features.

Hypothesis

Building a data-driven personal OS that runs 100% on the client — using in-browser ONNX Runtime neural models, vector search, and relational SQLite — delivers an engaging recruiter experience with zero server inference overhead and universal reusability.

Approach

Designed a decoupled architecture where structured JSON config seeds SQLite and generates vector embeddings. Created an OS desktop shell with window management, a hybrid BM25 + cosine similarity retrieval engine, and an in-browser local LLM assistant.

Implementation

Next.js App Router, React, TypeScript, Transformers.js / ONNX Runtime Web (WebGPU & WASM fallback), client-side SQLite with FTS5, and CSS-driven windowing mechanics.

Results

Fully responsive desktop OS running client-side with 0ms server inference costs, sub-millisecond 64D/384D hybrid vector retrieval, and universal plug-and-play profile configurability for any engineer.

What it taught me

• Zero hardcoded data guarantees clean extensibility and universal open-source reusability
• Client-side ONNX Runtime with WebGPU enables powerful in-browser neural assistance without backend GPU infrastructure
• Hybrid lexical (BM25) and semantic vector search eliminates hallucination while preserving high recall

Next

• Support pluggable web desktop themes and custom window widgets
• Expand local fine-tuned student LLMs with WebGPU quantization