Skill: Inbound Message Classifier

Purpose

Classify inbound LinkedIn messages within 15 minutes and route them appropriately: Class A (HOT) to Rick immediately, Class B (WARM) to Rick within 2 hours, Class C (NETWORKING) to accept/log, Class D (SPAM) to ignore, Class E (SUSPICIOUS) to alert human immediately.

What it does

•Analyzes sender profile (title, company, history)
•Parses message text for keywords, specificity, intent
•Applies deterministic classification rules
•Returns classification, confidence, routing instruction, and opportunity summary
•Logs every classification for audit trail

Why it matters

Recruiters with real opportunities are time-sensitive. A 15-minute response window can mean the difference between a qualified lead and a missed opportunity. Spam and scams waste time. Trust & Safety alerts require immediate action to protect account.

How to use it

Single message classification

python

from inbound_classifier.message_classifier import classify_message

result = classify_message({
    "message_id": "MSG-1234",
    "profile_id": "P-001",
    "sender_name": "John Smith",
    "sender_title": "Senior Recruiter",
    "sender_company": "TechStaff Inc",
    "message_text": "Hi! We have a Python/AWS role in Boston. Are you available?",
    "timestamp": "2026-02-15T10:30:00Z"
})

print(f"Classification: {result['class_label']}")
print(f"Route to: {result['routing']['route_to']}")
print(f"SLA: {result['routing']['sla_minutes']} minutes")

Batch classification

python

from inbound_classifier.message_classifier import classify_batch

results = classify_batch([msg1, msg2, msg3, ...])

# Results sorted by priority
for msg in results:
    if msg['routing']['priority'] in ['URGENT', 'CRITICAL']:
        print(f"ACTION REQUIRED: {msg['class_label']} from {msg['sender_context']['name']}")

CLI usage

bash

# Demo with sample messages
python message_classifier.py

# Single message from JSON
python message_classifier.py single_message.json

# Batch from JSON array
python message_classifier.py message_batch.json

Classification system

Class A: HOT (Urgent)

Indicators:

•Recruiter title (Technical Recruiter, Staffing Manager, etc.)
•AND specific tech mention OR availability question
•AND Tier 1 or Tier 2 staffing firm OR known good company

Confidence: 0.85-1.0

Example:

code

"Hi! We have a great Python/AWS role for a senior developer in Boston.
The client is looking for someone who can start immediately. Are you available?"

SLA: 15 minutes

Action: Route to Rick immediately with full sender and opportunity context

Class B: WARM (High)

Indicators:

•Recruiter title without strong indicators, OR
•HR title (but not dedicated recruiter), AND
•Opportunity mention (vague role, opportunity, project), OR
•Connection request with opportunity note

Confidence: 0.50-0.85

Example:

code

"Hi! I'd like to connect. We have some exciting tech opportunities
that might interest you based on your background."

SLA: 2 hours

Action: Queue for Rick evaluation within 2-hour window

Class C: NETWORKING (Low)

Indicators:

•Generic connection request, OR
•Industry peer / alumni / group member, OR
•No recruiter/HR signals, no opportunity mention

Confidence: 0.50-0.70

Example:

code

"Hey! Would love to connect with fellow Python developers.
Let's stay in touch and share industry insights."

SLA: None (no SLA)

Action: Accept and log. No escalation to Rick.

Class D: SPAM (Ignore)

Indicators:

•MLM keywords OR
•Insurance mention OR
•Generic sales pitch language, OR
•Mass InMail pattern

Confidence: 0.80-0.95

Example:

code

"Limited time offer! Get life insurance coverage today.
Don't miss out on this exclusive deal!"

SLA: None

Action: Ignore and log as spam

Class E: SUSPICIOUS (Critical)

Indicators:

•LinkedIn Trust & Safety language OR
•Verification/identity confirmation request OR
•Unusual activity warning OR
•Account security alert

Confidence: 1.0

Example:

code

"Verify your account identity immediately due to unusual activity
detected on your account. Click here to confirm."

SLA: 5 minutes (CRITICAL)

Action: Stop all activity immediately. Alert human and Z. Quarantine profile.

Input schema

json

{
  "message_id": "MSG-1234",
  "profile_id": "P-001",
  "sender_name": "John Smith",
  "sender_title": "Senior Technical Recruiter",
  "sender_company": "TechStaff Inc",
  "sender_id": "SEND-5678",
  "message_text": "Hi! We have a Python/AWS role...",
  "timestamp": "2026-02-15T10:30:00Z"
}

Field descriptions

•message_id: Unique identifier for the message
•profile_id: Which LinkedIn profile received the message
•sender_name: Name of the person sending message
•sender_title: Their job title
•sender_company: Their company
•sender_id: Their LinkedIn ID (optional)
•message_text: Full text of the message
•timestamp: When received (ISO-8601)

Output schema

json

{
  "message_id": "MSG-1234",
  "profile_id": "P-001",
  "classification": "A",
  "class_label": "HOT",
  "confidence": 0.92,
  "reasoning": [
    "recruiter_title",
    "specific_tech: Python, AWS",
    "availability_question",
    "tier1_firm"
  ],
  "routing": {
    "route_to": "rick",
    "priority": "URGENT",
    "sla_minutes": 15,
    "action": "evaluate"
  },
  "sender_context": {
    "name": "John Smith",
    "title": "Senior Technical Recruiter",
    "company": "TechStaff Inc",
    "sender_id": "SEND-5678",
    "company_tier": "TIER_1"
  },
  "opportunity_summary": "Python, AWS role, Boston",
  "timestamp": "2026-02-15T10:30:00Z",
  "classified_at": "2026-02-15T10:32:15Z"
}

Keyword lists

Recruiter titles

recruiter, technical recruiter, it recruiter, talent acquisition, staffing, headhunter, sourcerer, vendor, staffing manager, recruiting consultant, account manager, account executive

Specific tech/role keywords

Java, Python, C++, C#, JavaScript, TypeScript, React, Angular, Vue, Node.js, Django, Flask, Spring, Docker, Kubernetes, AWS, Azure, GCP, microservices, DevOps, SQL, NoSQL, MongoDB, role, position, opportunity, contract, full time, project

Availability keywords

available, interested, open to, looking for, seeking, exploring, immediate, asap, start date, notice period, rate, salary

Tier 1 firms

Cognizant, Infosys, TCS, Wipro, Accenture, Deloitte, PWC, KPMG, IBM, Microsoft, Google, Amazon, JPMorgan, Goldman Sachs

Tier 2 firms

Stripe, Figma, Databricks, Okta, Twilio, New Relic, Qualcomm, NVIDIA

MLM keywords

MLM, multi-level marketing, herbalife, amway, younique, monat, rodan, lularoe, itworks

Insurance keywords

Insurance, life insurance, health insurance, policy, premium, coverage, underwriting, allstate, state farm, progressive

Sales pitch keywords

Special offer, limited time, exclusive deal, click here, learn more, sign up now, buy now, free trial, money back, guarantee, earn money, work from home, passive income

Trust & Safety keywords

LinkedIn trust, LinkedIn safety, verification, confirm identity, verify account, unusual activity, unauthorized access, security alert

Classification algorithm

•Check for E (SUSPICIOUS) - If any Trust & Safety keyword found, return Class E immediately
•Check for D (SPAM) - If MLM, insurance, or sales pitch keywords found, return Class D
•
Check for A (HOT) - Score based on recruiter title + (tech OR availability) + (tier1 OR tier2)
- •If score >= 0.7: Return Class A
•
Check for B (WARM) - Score based on recruiter/HR + opportunity mention
- •If score >= 0.5: Return Class B
•Check for C (NETWORKING) - If any generic connection/peer/group signals found, return Class C
•Default to C - If no other match, classify as generic networking

Confidence scoring

•0.95-1.0: Very high confidence (recruiter + strong tech + availability + tier1 firm)
•0.85-0.94: High confidence (recruiter + tech/availability + tier firm)
•0.70-0.84: Medium-high confidence (recruiter with some indicators)
•0.50-0.69: Medium confidence (some recruiter signals or opportunity mention)
•< 0.50: Low confidence (generic message)

Opportunity extraction

For Class A and B messages, the classifier extracts:

•Tech stack: Matched keywords from message
•Location: Extracted geography if mentioned
•Role level: Implied from language (junior, mid, senior)
•Urgency: Based on keywords like "immediate", "asap", "urgent"

Batch processing

When classifying multiple messages:

•Sort results by priority (E > A > B > C > D)
•Within same priority, sort by confidence (descending)
•This order tells you what to handle first

Performance

•Single message: <50ms
•Batch of 100: <5s
•Memory efficient, no database calls

Error handling

•Missing sender_title: Treated as non-recruiter
•Missing message_text: Classified as low-confidence
•Empty fields: Skipped without error
•Invalid timestamp: Uses current time

Integration points

•Inbound Monitor: Receives new messages and sends to classifier
•Leroy Core: Reviews classifications, routes to Rick or alerts
•Rick: Receives Class A/B leads within SLA
•Human/EM: Receives Class E alerts immediately
•Logging: Every classification appended to history

SLA enforcement

•Class A (HOT): Must route to Rick within 15 minutes
•Class B (WARM): Must route to Rick within 2 hours
•Class E (SUSPICIOUS): Must alert human within 5 minutes
•Class C/D: No SLA, process in batch

Future enhancements

•Machine learning scoring (retrain on historical classifications)
•Reputation tracking (build history per sender)
•Budget/rate mention detection
•Contract length mention detection
•Client confidentiality keywords