Gyrus

An agentic browser that tries to make you a sharper researcher instead of just outsourcing the thinking to an agent.

Built in 36-ish hours at WeaveHacks — Agent Protocols Hackathon (Weights & Biases + Google), San Francisco (July 2025).

Gyrus started life as Clay — we rebranded for the hackathon in case Weights & Biases minded the name. The research behind it later evolved into Friction, the same question taken to the OS and hardware level.

Demo_Gyrus.mov

What it does

Gyrus watches what you search and builds a memory graph of it in Neo4j: queries, the concepts they cluster into, and the links you click. Every query goes through an intent classifier first — Research, Answer, Transactional, News, or Navigational — and what happens next depends on the intent. A Research or News query gets handed to a CrewAI "crew" that pulls from Exa, arXiv, Semantic Scholar, GDELT, or NewsAPI and comes back with a consolidated set of links. Everything else just passes through.

Before any query text leaves the machine for an external API or model, a PII scrubber (CHAAP) strips emails, phone numbers, SSNs, card numbers, and names out of it.

The browser itself is an Electron app with a custom webview shell, not a wrapper around Chrome's UI.

How it works

• Intent detection — backend/src/fivedvector.py scores each query against two signal sources: a lookup over past DB behavior, and a zero-shot DeBERTa classifier (backend/tools/intent_zero_shot_classifier.py, ensembled over four hypothesis templates) — and takes whichever source is most confident. model-finetune/roberta.py is a separate Colab notebook that fine-tunes RoBERTa on the ORCAS-I-2M query dataset (~2M queries) for the same intent split, to check the zero-shot approach against a supervised baseline.
• Memory graph — backend/src/query_orch.py embeds each query's concept with sentence-transformers (MiniLM) and writes it into Neo4j as Concept, Query, and Link nodes, connected by SEARCHED_BY and CLICKED edges. A new concept merges into an existing one above a cosine-similarity threshold instead of duplicating it.
• Crews — backend/MCP/newscrew_http.py and researchcrew.py are CrewAI agents wired to Exa, arXiv, Semantic Scholar, GDELT, and NewsAPI through an MCP server, with Weave tracking on the CrewAI runs.
• Privacy — backend/tools/chaap_anonymize.py is a regex-based obfuscator that scrubs PII before a query touches any external API or model.
• Frontend — browser/ is an Electron app (main.js, preload.js) talking to the Flask backend (backend/src/app.py) over REST.

Prototype

prototype/crew_routing.py draws a concept sketch of the shipped routing architecture — no invented numbers, just the pipeline as built: a browsing signal (history, current tabs, queries) passes through the CHAAP PII obfuscation layer, then feeds both the Neo4j memory graph and the zero-shot intent classifier, which together settle on an intent (research / study / shop / doomscroll). That intent picks a matching CrewAI crew, which pulls from arXiv, Semantic Scholar, Exa, GDELT, and NewsAPI, and lands on an assist action (surface papers, quiz, block-the-feed nudge).

Run it locally to regenerate the figure:

MPLCONFIGDIR=/home/arya/projects/hackathons/.mplcache /home/arya/projects/hackathons/.venv/bin/python prototype/crew_routing.py

Team

• Arya Garg — github.com/Aryagarg23 — browser frontend (Electron UI, styling), backend wiring.
• Kaaustaaub Shankar — memory graph, API.
• Raihan Rafeek — PII obfuscation (CHAAP), memory-intent analysis.

Links

• Devpost
• Writeup
• aryagarg23.com
• Devpost profile

More hackathon builds

• WhiteBox — traceable GraphRAG over medical literature (Future of Data 2024, 1st place)
• G-Code-Assembler — G-code assembly + STL visualization (MakeUC 2024, Kinetic Vision winner)
• Terminally-Addicted — Spotify, GitHub, GPT and YouTube without leaving the terminal (HackOHI/O 2024)
• Memento — digital memory journal for Alzheimer's patients and caregivers (RevolutionUC 2024, 3rd overall)
• Buycott — barcode scan -> parent company -> NLP stance on social issues (MakeUC 2023, 1st overall)
• SignLink — video calls with real-time ASL fingerspelling to text (BoilerMake X 2023)
• Kuka Arm Viz — interactive 7-DOF robot arm in WebGL with inverse kinematics (RevolutionUC 2023)
• Hi-Five — anonymous friend-matching on OCEAN personality vectors (SASEhack 2024)
• Friction — speculative OS + hardware that protects flow state with physical friction (Fig Build 2026)