👔 HyredLab: Agentic Career Orchestration Platform
HyredLab is an enterprise-grade, LLM-driven job application tracker that leverages Agentic Workflows and Chain-of-Thought (CoT) reasoning to autonomously synchronize your career search. It replaces manual data entry with a multi-agent system that ingests, cleans, and structures unstructured data from Gmail.
Architecture Deep Dive 🏗️
Live: www.hyredlab.com
🖼️ Visual Tour
Dashboard Overview
A mission-control interface featuring real-time telemetry, Optimistic UI updates, and Semantic Search capabilities.
Autonomous Ingestion Engine
The Neural Batch Pipeline processes high-volume unstructured email streams using Concurrent Rate-Limited Queues, bypassing serverless timeouts via client-side orchestration.
Semantic Identity Resolution
HyredLab uses L4 Semantic Matching (a proprietary ranking algorithm) to merge fragmented signals (updates, interviews) into a single Canonical Application Entity.
Neural Decision Logs
Full observability into the LLM's reasoning trace. Audit the Chain-of-Thought steps as the agent evaluates candidates and resolves edge cases.
Authorized Access Control
Enterprise-ready RBAC (Role-Based Access Control) pattern ensuring data sovereignty and tenant isolation.
Analytics & Insights
Data visualization layer using Server-Side Aggregation to compute funnel conversion rates and activity velocity metrics.
Pipeline Board (Kanban)
Interactive state management system featuring dnd-kit integration and Optimistic UI patterns. Now integrated into the main Dashboard for seamless switching.
Advanced Grid View
A high-density Tabular Grid with Column Selection capabilities. Toggle visibility for Salary, Recruiter, Hiring Manager, and AI-Extracted Rejection Reasons.
AI Bulk Import
ETL Pipeline for unstructured text. Uses Few-Shot Prompting to parse legacy data formats (Excel, LinkedIn) into standardized schemas.
✨ Key Capabilities
- 🤖 Autonomous Ingestion Agent: A background worker that polls Gmail APIs using Incremental Sync strategies.
- 🧠 LLM Entity Extraction: Uses Zero-Shot Extraction to parse multi-modal email content into structured JSON schemas.
- 🔍 RAG-Enhanced Reflexion: A Self-Healing Module that critiques its own outputs against a vector database of "Learned Rules" to correct hallucinations.
- 📂 Semantic Threading: Uses Fuzzy Logic and Time-Series Analysis to cluster related events into a unified lifecycle.
- 🛠️ Self-Maintaining Database: Automated cron jobs for Schema Normalization and orphan data cleanup.
- 📊 Analytics Dashboard: Comprehensive funnel visualization (Sankey) and activity tracking with Dark Mode Optimization.
- 📋 Unified Pipeline: Switch between Card, Grid, and Kanban views from a centralized mission control.
- 📥 AI Bulk Import: Parse unstructured text lists (Excel, LinkedIn) into structured job data.
- 🚫 Rejection Extraction: Automatically identifies and summarizes rejection reasons from candidate communications.
- ⚙️ Column Customization: Tailor the Grid view to show exactly what you need (Salary, Hiring Manager, etc.).
- 🛠️ Maintenance Tools: Built-in controls for database consolidation and cleanup.
- 🔒 Privacy First: Gmail is read with the
gmail.readonly scope, and your data lives in your own PostgreSQL database (not a local SQLite file).
- 🚀 Vercel Analytics: Built-in performance and traffic tracking via
@vercel/analytics.
🚀 Tech Stack
🧠 Agentic AI Flow
The core of HyredLab is a multi-stage agentic workflow designed to handle the noise of a real inbox:
-
Extraction Agent (Front-Line):
- Scans incoming emails and structures unstructured text into a standard Job schema.
- Uses Self-Correction Logic: It fetches "Learned Rules" from the database (rules established by your previous "Ignore" or "Correction" actions) to avoid repeating mistakes like classifying newsletters as applications.
-
The AI Judge (Identity Resolution):
- When new data arrives, a context-aware agent determines if it belongs to an existing application or is a fresh one.
- It performs Semantic Reasoning (understanding that "SDE" at Google and "Software Dev" at Alphabet might be the same job) and Time-Based Logic (distinguishing a status update from a re-application months later).
-
Reflexion Loop (Self-Healing):
- Triggered when you request a "Deep Refresh". The AI re-analyzes original emails plus any follow-up threads.
- It performs Critique and Correction: Identifying why the previous attempt was incorrect and using your direct feedback as the ultimate ground truth.
-
Maintenance Agent (Consolidation):
- Automatically runs across your database to merge legacy duplicates or fragmented threads into a cohesive career timeline.
-
Neural Batch Sync (High-Performance Pipeline):
- Multi-Stage Orchestration: Splits the sync into a Prepare stage (metadata discovery) and a Process stage (deep analysis), allowing it to handle up to 365 days of history.
- Parallel Batching: Uses a client-orchestrated parallel threading model to bypass Vercel serverless timeout limits. It processes emails in concurrent batches of 25, delivering high throughput while maintaining platform stability.
- Surgical Search: Supports advanced date-range parameters (
after: and before:) mapped directly to the Gmail API query engine for precise historical discovery.
🛠️ Getting Started
1. Prerequisites
- Node.js 18+
- A Google Cloud Project (for Gmail/OAuth)
- A Gemini API Key (or OpenRouter)
2. Installation
git clone https://github.com/laadtushar/job-hunt-dashboard.git
cd job-hunt-dashboard
npm install
3. Environment Setup
Create a .env file (or .env.local) in the project root. The repo ships no .env.example, and .env* is git-ignored, so never commit it. Set:
POSTGRES_PRISMA_URL & POSTGRES_URL_NON_POOLING: Your PostgreSQL connection strings (pooled and direct). The database needs the pgvector extension. These are the names Vercel's Postgres integration provides.GOOGLE_CLIENT_ID & GOOGLE_CLIENT_SECRET: From your GCP Console.AUTH_SECRET: Generate one using npx auth secret.GEMINI_API_KEY: From Google AI Studio. Used when AI_PROVIDER is unset or GEMINI.SUPERADMIN_EMAIL: The Google account that is always allowed in and gets the admin role. Without it, the first user to sign in is allowed.
Optional:
AI_PROVIDER: GEMINI (default) or OPENROUTER. With OPENROUTER, also set OPENROUTER_API_KEY (and OPENROUTER_MODEL to pick the model).EMBEDDING_PROVIDER: GEMINI, OPENROUTER or HUGGINGFACE (runs a local model). Defaults to AI_PROVIDER, then GEMINI. OPENROUTER_EMBEDDING_MODEL picks the OpenRouter embedding model.NEXT_PUBLIC_LOGO_DEV_TOKEN: A logo.dev token for company logos.
4. Database Setup
npx prisma generate
npx prisma migrate dev
5. Launch
npm run dev
📖 Usage Tips
- Syncing: Click the Sync button on the dashboard to pull latest emails.
- Deep Re-analysis: If a job info looks wrong, use the "Refresh" icon on the job card. You can provide a hint (e.g., "I actually got rejected yesterday") to guide the AI's deep search.
- Ignore: Use the trash icon for emails that aren't actually jobs; this helps the AI learn.
📄 License
MIT License - feel free to use and modify for your own job hunt!
