social knowledge graph UI
grok 3 mini search feature
find and reach out to top candidates
autogenerate a DM on X
GrokRuit pipeline

Grokruit: Discover the Universe of Talent

Try it live: https://grok-underrated-recruiter.vercel.app/

Sourcing talent is the most critical problem that xAI has not been able to fully grok: finding the underrated, high-signal engineers. Getting from 100K candidates to 10 requires human recruiters to search, rank, and evaluate one by one. We reverse-engineered Elon Musk's early hire strategies to build Grokruit.

Inspiration

When discussing with Christina and xAI engineers, it was brought up that sourcing is the primary bottleneck. We believe this tool can automate recruitment pipelines while providing high-signal talents.

What It Does

Grokruit addresses this challenge using PageRank + Grok as LLM judge to find underrated candidates.

Human recruiters use explore (discover candidates through existing connections) and exploit (in-depth analysis on top K candidates). Grokruit automates the process using a 4-stage pipeline:

1. Social Knowledge Graph Construction

Start with initial seed accounts (e.g., xAI engineers)
Expand via follows, retweets, replies to build interaction edges with weighted signals
Edge weights: Following (5.0), Retweets (3.0), Replies (2.5), Likes (1.0)
Result: 21K+ nodes, 28K+ edges

2. Fast LLM Screen (grok-4-1-fast-non-reasoning)

Screen bios for technical relevance
Filter out xAI/X employees (already hired), organization accounts, non-technical accounts
First-pass to identify engineers
16K → 500 candidates in minutes (~45% filter rate)

3. PageRank + Learnable Score

Personalized PageRank with seeds as the personalization vector
Underratedness Score: PageRank / log(followers) — boosts low-profile but high-potential candidates who may be missed by recruiters
PageRank has learnable components which can be updated with online feedback from recruiters

4. Deep Evaluation (grok-4-1-fast-reasoning + xAI Search Tools)

Autonomous information gathering via web_search and x_search tools
Searches GitHub, LinkedIn, and X posts for each candidate
5-criterion rubric (0-100 scale):
- Technical Depth (25%)
- Project Evidence (25%)
- Mission Alignment (20%)
- Exceptional Ability (20%)
- Communication (10%)

Key Features

Login with X (OAuth 2.0): Authenticate with your X account to unlock personalized features like saving candidates and sending DMs.
Natural Language Search with Thinking Trace: Search for candidates using natural language queries (e.g., "ML engineers with robotics experience"). Watch Grok's reasoning process in real-time as it ranks and filters candidates.
Interactive Knowledge Graph Visualization: Explore the social network graph with force-directed visualization. Filter by PageRank, depth, and relevance. See how candidates are connected through the network.
AI-Powered Personalized DM Generation Generate personalized outreach messages using Grok. Choose your tone (professional, casual, enthusiastic), add custom context, and preview the message before sending.
Send DMs via X API: Actually send DMs directly through the app using your authenticated X account. No more copy-pasting!
Real-Time Handle Submission: Submit X handles for evaluation and watch the pipeline process them in real-time. Track status from submission → fast screen → deep evaluation → completion.
Expand the Knowledge Graph: Admins can add new seed accounts to expand the knowledge graph. The system automatically crawls their network and discovers new candidates.
Save & Track Candidates: Build your candidate shortlist. Save interesting profiles for later review and outreach.
Admin Panel: Approve/reject handle submissions, manage the evaluation queue, and monitor system health.

How We Built It

Backend

X API (OAuth1 + Bearer Token) for graph crawling
xAI API with multiple Grok models:
- grok-4-1-fast-non-reasoning for fast screening
- grok-4-1-fast-reasoning + tool calls for deep evaluation
- grok-3-mini for natural language search with thinking trace
xAI SDK with web_search() and x_search() tools
NetworkX for PageRank computation
FastAPI with SSE streaming for real-time updates
SQLite for saved candidates, DM history, and submissions
Redis for OAuth state and job queuing (arq)

Frontend

React/TypeScript with Vite
react-force-graph-2d for graph visualization
React Router for navigation
Real-time streaming UI for search and DM generation

Deployment

Fly.io for backend with persistent volumes
Vercel for frontend
Redis Cloud for distributed state

Challenges We Ran Into

X API rate limits: Aggressive caching at every layer (X responses, Grok evaluations, graph state) to minimize API calls
High volume of noisy accounts: Use Grok fast 4.1 as first-pass to aggressively filter down candidates before expensive deep evaluation
Making the tool usable: Building an intuitive UI for social graphs, streaming search results, and ensuring query results match recruiter expectations
OAuth token management: Storing and managing X access tokens securely for DM sending functionality

Accomplishments We're Proud Of

We worked hard to make something immediately useful and deployable:

End-to-end automation: From X handle to scored candidate profile with GitHub/LinkedIn links
Real-time graph expansion: Add new sources and run screening directly from the UI
AI-generated personalized DMs: Using candidate evaluation data to craft relevant outreach
Direct DM sending: Actually send messages through X API, not just copy-paste
Live streaming UI: Watch Grok think through candidate rankings in real-time

What's Next for Grokruit

Fine-tune LLM adapter: Learn from recruiter feedback (whether or not the candidate made it through, +/-1 reward)
Interview preparation: Integrate with interviewer prep and automate the rest of the recruiting pipeline
Synthetic dataset evaluation: Build benchmark datasets to measure system performance
Expand to other platforms: LinkedIn, GitHub, Discord communities