sanyam.
browser · nothing leaves the devicetab 1tab 2tab 3capturecontent scriptembeddingsTransformers.jssession detectproject inference3-layer searchkeyword, vector, rerankknowledge graphONNX WASM84% clustering precision · NDCG@5 0.91 · F1 0.88
capturecontent scriptembed on the deviceTransformers.js, ONNXsessionsproject inference3-layer searchkeyword, vector, rerankknowledge graphnothing leaves the deviceNDCG@5 0.91, 84% clustering precision.
All projects

KONTA

Context-Aware Browsing Extension

Browsers forget. Close a tab and the research trail is gone. KONTA captures every session as you browse and organizes it into a searchable knowledge graph, so the work you did last week is still there when you need it.

Result
84% clustering precision
NDCG@5: 0.91
F1: 0.88 detection
What it does

Chrome extension that turns your browsing into a personal knowledge base. It detects sessions automatically, infers which project each page belongs to, and makes all of it searchable, with every model running inside the browser.

What I built
  • Three-layer retrieval: strict keyword matching first, semantic embeddings second, an ML re-ranker on top. Rank@1 and NDCG@5 are measured by automated evaluation scripts in the repo.
  • Interactive knowledge graph built with Louvain community detection, validated against 8 topic clusters plus deliberately ambiguous noise pages.
How it works
  1. 01Chrome Extension
  2. 02Background Workers (session detection + ML ranking + embedding engine)
  3. 03Content Scripts (page capture + Google Search integration)
  4. 04Knowledge Graph UI
Features
  • Automatic session capture with ML-based project inference across open tabs
  • Three-layer search: BM25 keyword, Transformers.js semantic embeddings, then an ML re-ranker
  • Force-directed knowledge graph with Louvain community detection, rendered with react-force-graph
  • Google Search integration that surfaces your own past research next to live results
  • Focus mode, context-aware reminders, and browser-history import to backfill the graph
  • All inference runs on-device through Transformers.js and ONNX WASM; nothing leaves the browser