AI Job Connector Agent

The AI Job Connector Agent automates job search outreach by intelligently finding HR/recruiter contacts for specified job postings and generating personalised connection messages.

Comment

Inspiration

Job hunting is one of the most time-consuming and emotionally draining processes professionals face. After watching friends spend countless hours researching companies, finding the right people to contact, and crafting personalised messages—only to receive no response—we realised there had to be a better way.

The inspiration struck when we saw how much time job seekers waste on repetitive tasks: manually searching LinkedIn for recruiters, copying job descriptions, crafting individual messages for each application, and keeping track of dozens of conversations. What if we could automate 90% of this work while making the outreach more personal and effective?

We envisioned an AI agent that doesn't just help you apply to jobs—it becomes your intelligent networking assistant, finding the right people, understanding your unique value proposition, and crafting authentic, personalised messages that actually get responses.

What it does

AI Job Connector Agent is an intelligent automation system that transforms job hunting from a tedious manual process into an efficient, AI-powered workflow:

Core Features

  1. Smart HR Contact Discovery

    • Automatically searches for recruiters and hiring managers at target companies
    • Uses advanced web search (Tavily API) to find LinkedIn profiles
    • Identifies the most relevant contacts based on job title and company
    • Returns confidence scores and profile URLs

  2. CV Intelligence

    • Parses your resume (PDF, DOCX, or TXT) using natural language processing
    • Extracts key skills, experience, education, and achievements
    • Creates a semantic understanding of your professional profile
    • Uses this context to personalise every message

  3. Personalized Message Generation

    • Leverages AWS Bedrock with Claude 3.5 Sonnet to craft authentic outreach messages
    • Adapts tone (professional, friendly, enthusiastic) based on your preference
    • Optimises for channel (LinkedIn, email, or generic)
    • Generates concise, medium, or detailed messages as needed
    • Highlights relevant experience that matches the specific job

  4. Batch Processing

    • Process multiple job applications simultaneously
    • Upload JSON files with dozens of job postings
    • Track progress in real-time with live updates
    • Download results as CSV or JSON for easy integration with your workflow

  5. Production-Ready Web UI

    • Beautiful, intuitive Streamlit interface
    • Drag-and-drop file uploads
    • Real-time progress tracking
    • Copy-paste ready messages
    • Mobile-responsive design

The Result?

What used to take 30-60 minutes per job application now takes seconds. And the messages aren't generic templates—they're intelligently crafted, genuinely personalised outreach that reads like you wrote it yourself.

How we built it

Technology Stack

AI & LLM:

  • AWS Bedrock with Claude 3.5 Sonnet (anthropic. claude-3.5 Sonnet-20240620)
  • Cross-region inference profiles for optimal performance in the EU
  • Tool use/function calling for agentic behaviour
  • Streaming responses for real-time feedback

Backend:

  • Python 3.11+ for modern async/await patterns
  • Pydantic v2 for robust data validation and serialisation
  • boto3 for AWS SDK integration
  • Custom agent runtime with tool orchestration
  • Exponential backoff retry logic for API resilience

Search & Discovery:

  • Tavily API for AI-optimized web search
  • Serper API as fallback option
  • Focused domain filtering (LinkedIn, Indeed, Glassdoor)
  • Result scoring and ranking

Infrastructure:

  • AWS ECS Fargate for serverless container hosting
  • Application Load Balancer for traffic distribution
  • Amazon ECR for Docker image registry
  • AWS CDK (Python) for infrastructure as code
  • CloudWatch Logs for monitoring and debugging

Frontend:

  • Streamlit for rapid UI development
  • Real-time progress bars and status updates
  • File upload handling (CV parsing, JSON batch import)
  • Data export (CSV, JSON formats)

Development Process

  1. Agent Design - Designed the agentic architecture with tool use patterns
  2. Core Runtime - Built the AgentRuntime class to handle Bedrock invocations
  3. Tool Integration - Implemented HR Lookup tool with Tavily search
  4. Message Generation - Created message generation service with prompt engineering
  5. Web Interface - Developed Streamlit UI with batch processing
  6. Infrastructure - Deployed to AWS ECS Fargate with CDK
  7. Production Hardening - Added retry logic, error handling, rate limiting

Key Design Decisions

Why Bedrock + Claude?

  • Most advanced reasoning capabilities for understanding context
  • Excellent at following complex instructions
  • Natural, human-like message generation
  • Native tool use support

Why Tavily over Google?

  • Optimised for AI agents (cleaner results, better structure)
  • Advanced search depth options
  • Domain filtering built-in
  • Faster response times

Why ECS Fargate?

  • Serverless container management (no EC2 to manage)
  • Auto-scaling based on demand
  • Pay only for what you use
  • Easy integration with other AWS services

Challenges we ran into

Building production AI systems is never smooth sailing. Here are the real challenges we faced:

Bedrock API Throttling

The Problem: When processing multiple jobs in batch, we hit AWS Bedrock's rate limits hard. The API returned:

ThrottlingException: Too many requests, please wait before trying again

The Solution: Implemented exponential backoff retry logic with jitter:

  • Base delay: 2 seconds
  • Exponential increase: 2^attempt × base_delay
  • Max retries: 5 attempts
  • Random jitter to avoid thundering herd
  • Graceful degradation with error messages

This reduced the failure rate from ~40% to <1% in batch processing.

Message Quality Control

The Problem: Early messages were either too generic ("I'm interested in this role") or too aggressive ("I'm the perfect candidate").

The Solution: Extensive prompt engineering:

  • System prompts emphasising authenticity
  • Examples of good vs. bad messages
  • Tone calibration with temperature tuning
  • Multiple revision cycles with user feedback
  • Configurable tone/length/channel parameters

Accomplishments that we're proud of

Technical Achievements

  1. Production-Grade Reliability

    • Exponential backoff handles 99%+ of transient failures
    • Graceful degradation when services are unavailable
    • Comprehensive error messages guide users to solutions

  2. Real Agentic Behavior

    • Claude makes autonomous decisions about which contacts to prioritise
    • Tool use loop handles multi-step reasoning
    • Dynamic prompt construction based on context

  3. Sub-Second Response Times

    • Optimized Bedrock invocations with streaming
    • Parallel processing where possible
    • Efficient CV parsing with caching

  4. Scalable Architecture

    • Can process 100+ jobs in a single batch
    • Auto-scales based on load
    • Stateless design for horizontal scaling

User Experience Wins

  1. Zero Configuration Required

    • No API keys to enter (pre-configured)
    • No complex setup or installation
    • Works immediately after visiting the URL

  2. Beautiful, Intuitive UI

    • Drag-and-drop file uploads
    • Real-time progress tracking
    • Copy-paste ready messages
    • Export results for integration with your workflow

  3. Genuine Personalization

    • Messages reference specific experience from your CV
    • Tone adapts to company culture (startup vs. enterprise)
    • Highlights the most relevant skills for each role

Innovation Highlights

  1. CV Intelligence

    • Semantic understanding of your career story
    • Automatic skill extraction and categorisation
    • Context-aware message personalisation

  2. Smart Contact Discovery

    • Finds the right person, not just any recruiter
    • Confidence scoring helps you prioritise
    • LinkedIn profile URLs for direct outreach

  3. Agentic Tool Use

    • Claude decides when to search vs. when to respond
    • Multi-turn reasoning for complex scenarios
    • Self-correction when searches return no results

What we learned

Technical Insights

  1. AWS Bedrock Nuances

    • Cross-region inference profiles are essential for EU deployments
    • Throttling limits are real—always implement retry logic
    • Streaming responses dramatically improve perceived performance
    • Tool use requires careful prompt engineering for reliability

  2. Agent Design Patterns

    • Always provide tools with clear, unambiguous descriptions
    • Return structured data (JSON), not free text
    • Implement tool result validation
    • Log everything—debugging agentic behaviour is hard without visibility

  3. Production AI Challenges

    • Rate limiting is the #1 issue with production LLM apps
    • Error handling is 50% of the code in production systems
    • User feedback loops are essential for prompt improvement
    • Cost monitoring is critical (each Claude call costs money)

  4. Infrastructure as Code

    • AWS CDK makes infrastructure predictable and repeatable
    • Docker multi-stage builds are worth the complexity
    • Environment variable management is harder than it looks
    • Always include health checks in production deployments

Product Insights

  1. AI Agents Need Guardrails

    • Without retry logic, agents fail unpredictably
    • Tool execution must be idempotent
    • Users need visibility into what the agent is doing
    • Errors must be actionable, not cryptic

  2. UX Matters for AI Products

    • Users don't care about the technology—they care about results
    • Real-time progress indicators build trust
    • Copy-paste-ready outputs save time
    • Export formats matter (CSV for Excel, JSON for developers)

  3. Personalization is King

    • Generic messages get ignored
    • Specific references to experience get responses
    • Tone matching is more important than content
    • Authenticity beats perfection

Personal Growth

  • Cloud Architecture: Learned to design scalable, serverless systems
  • Agent Development: Understood the nuances of tool use and multi-turn reasoning
  • Prompt Engineering: Discovered how small prompt changes dramatically affect output
  • Production Operations: Experienced the challenges of running AI systems at scale
  • Problem Solving: Developed debugging skills for non-deterministic systems

What's next for AI Job Connector Agent

We have ambitious plans to make the AI Job Connector Agent even more powerful:

Near-Term Improvements (Next 2-4 Weeks)

  1. Email Integration

    • Direct send via AWS SES
    • Email tracking (opens, clicks)
    • Automated follow-ups
    • Response parsing

  2. LinkedIn API Integration

    • Direct message sending
    • Connection request automation
    • Profile view tracking
    • InMail credits management

  3. Advanced CV Analysis

    • Gap detection and recommendations
    • Skill matching with job requirements
    • Career progression analysis
    • Salary range suggestions

  4. Enhanced Search

    • GitHub profile integration
    • Company culture analysis
    • Funding/growth stage detection
    • Tech stack matching

  5. Analytics Dashboard

    • Response rate tracking
    • A/B testing for message variations
    • Best performing templates
    • Time-to-response metrics

Medium-Term Features (1-3 Months)

  1. Multi-Language Support

    • Translate messages to the target language
    • Cultural adaptation (US vs. EU vs. Asia)
    • Local idiom usage
    • Region-specific best practices

  2. CRM Integration

    • Salesforce connector
    • HubSpot integration
    • Google Sheets sync
    • Notion database export

  3. Advanced Personalization

    • Company research summaries
    • News mention integration
    • Social media activity analysis
    • Mutual connection detection

  4. Interview Preparation

    • Question generation based on job description
    • Answer suggestions based on CV
    • Mock interview scenarios
    • Weakness mitigation strategies

  5. Mobile App

    • iOS and Android native apps
    • Push notifications for responses
    • Quick message editing on the go
    • Voice-to-text for CV updates

Long-Term Vision (3-12 Months)

  1. AI Interview Coach

    • Real-time interview feedback
    • Body language analysis (video)
    • Speech pattern optimisation
    • Confidence scoring

  2. Career Path Advisor

    • Multi-year career planning
    • Skill gap identification
    • Course recommendations
    • Salary progression forecasting

  3. Marketplace Integration

    • Direct application to job boards
    • ATS integration (Greenhouse, Lever, Workday)
    • Recruiter matching platform
    • Freelance project finder

  4. Team Features

    • Shared contact databases
    • Collaborative message editing
    • Team analytics and leaderboards
    • Enterprise SSO and permissions

  5. AI Agent Marketplace

    • Custom agent creation platform
    • Share and monetise agent templates
    • Integration ecosystem
    • Third-party tool connectors

Built With

Share this project:

Updates