AI-native Predictive GTM CopilotPinkMate — Agentic Predictive GTM Copilot
Turn raw leads into a prioritised, outreach-ready pipeline using enrichment signals, predictive ML scoring, and agentic workflow orchestration.
Under active development. Public MVP documentation and technical repository are available below.
Built by Karan Sehgal (ex-Amex, WegoPro, Nokia) — founder-led AI GTM system combining ML scoring + agentic orchestration.
Built With Production-Grade Tools
Python • Scikit-learn • Pandas • NumPy
LangChain • LangGraph • LangSmith
Custom workflow engine • State management
CSV • LinkedIn Sales Navigator • Apollo.io
How It Works
System Components
Firmographics + buyer intent signals as model inputs.
Interpretable ML baselines for prioritisation. The MVP combines classical ML models (Naive Bayes, Random Forest, Gradient Boosting) with agent-orchestrated workflows using LangChain and LangGraph to predict conversion likelihood, generate ICP tiers, and trigger GTM actions.
ICP fit + readiness tiers drive routing.
Workflow transitions and triggers to next actions. Autonomous agents monitor segment changes, execute multi-step sequences (enrichment → scoring → routing → outreach), handle conditional branching, and learn from outcomes to optimize future decisioning paths.
Multi-Agent Pipeline Design
Six specialized agents working in orchestrated harmony
Orchestration note: Agents communicate via LangGraph state machines with error handling, retries, and conditional routing. Vector-based memory stores account history and interaction context for personalization.
Predictive Scoring Under the Hood
Naive Bayes
Baseline probabilistic classifier for quick lead categorization
Random Forest
Feature importance extraction and non-linear relationship detection
Gradient Boosting (XGBoost/LightGBM)
Primary scoring model with high predictive accuracy
The system extracts and engineers features across multiple dimensions:
- Firmographic signals: Company size, industry, revenue band, growth stage
- Behavioral signals: Website visits, content engagement, email opens
- Intent signals: Job postings, technology stack changes, funding events
- Temporal patterns: Engagement velocity, time-since-last-action
Each lead receives:
- Conversion probability score (0-1) with confidence intervals
- ICP fit tier (High/Medium/Low)
- Feature attribution showing which signals drove the score
- Low-certainty flags for manual review
- Structured response format for downstream agent consumption
A High-Performance Growth Engine Built on ML + Agentic Intelligence
From Static Lists to Autonomous GTM Execution
Traditional CRMs are storage systems. They hold lead data and depend entirely on manual prioritization, static segmentation rules, and human-driven follow-up. PinkMate is fundamentally different — it's a decisioning and execution layer that combines predictive ML with agentic orchestration.
The Result:
A self-improving GTM system where enrichment feeds prediction, prediction drives segmentation, and segmentation triggers execution — all without manual intervention.
MVP Constraints & Known Limitations
PinkMate is an early-stage MVP under active development. Current limitations include:
- Processing Mode:Batch processing only — no real-time API yet. Scoring runs are triggered manually or on schedule.
- Use Case Scope:Optimized for B2B SaaS outbound initially. Other verticals (e-commerce, services) not yet tested or tuned.
- Enrichment Setup:Manual connector configuration required. Automated enrichment source discovery is planned for Phase 2.
- Collaboration:Single-user mode. Team-based workflows, role permissions, and shared dashboards are in development.
- Model Maturity:Baseline models trained on synthetic and limited real-world data. Performance improves as engagement signals accumulate.
These constraints are being addressed systematically in the development roadmap. Feedback from early users directly shapes prioritization.
Use Cases
Outbound for B2B SaaS teams
ABM and ICP tiering
Predictive lead qualification
Pipeline prioritisation and sequencing
What You Get
- Prioritised lead list with readiness tiers
- Enriched records with key firmographic fields
- Scoring rationale (MVP-level transparency)
- Suggested next-best action trigger per segment
- Public build artefacts (repo + documentation)
System Architecture Overview
Lead Sources
CSV, LinkedIn, Apollo
Enrichment Layer
Firmographic + intent data
Feature Engineering
Signal extraction
ML Scoring Engine
NB, RF, GBM models
Outreach Execution
Email/LinkedIn triggers
Agentic Router
LangGraph state machine
Segmentation Logic
Readiness + ICP routing
ICP Classifier
High/Med/Low tiers
Outcome Tracking
Meeting booked, opp created
Feedback Loop
Label collection
Model Retraining
Periodic updates
Verification & Build Evidence
Technical implementation, ML workflows, and agentic orchestration logic.
View GitHub RepositoryProduct spec, workflow definitions, and feature roadmap.
View Documentation- agents/ - Six specialized agent implementations
- workflows/ - LangGraph orchestration and routing logic
- models/ - ML model artifacts and configuration files
- docs/ - Architecture documentation and workflow diagrams
This is an early MVP and technical proof-of-concept under active development. The landing page, documentation, and repository reflect the current build state.
Frequently Asked Questions
Built by a Founder Who Ships
I'm Karan Sehgal, Founder of PinkMate. I'm designing the GTM workflow, defining the predictive scoring approach (Python-based ML baselines + feature design), and shaping the agentic execution model that moves GTM from manual list-sorting to AI-native prioritisation and action.
Background: Previously at American Express (marketing analytics), WegoPro (growth engineering), and intern at Nokia (network engineering). Building in public — all code, documentation, and development progress is open and verifiable.
Active Development Phases
- ✅Core ML scoring pipeline
- ✅LangChain + LangGraph orchestration
- ✅CSV data ingestion
- ✅Basic segmentation logic
- ✅Public documentation + GitHub
- 🔄Feature engineering optimization
- 🔄Model evaluation framework
- ⏳Expand enrichment sources
- ⏳Behavioral signal tracking
- ⏳Feedback loop implementation
- ⏳Enhanced agentic decision trees
- ⏳A/B testing framework
- ⏳LLM-enhanced scoring layers
- ⏳RAG-based context grounding
- ⏳Autonomous outreach sequencing
- ⏳CRM integrations (HubSpot, Salesforce)
- ⏳Real-time scoring API
- ⏳Self-service configuration UI
Recent Updates
- Added Gradient Boosting models to scoring ensemble
- Implemented LangGraph state machine for workflow orchestration
- Expanded feature engineering pipeline (temporal patterns + intent signals)
- Published public documentation + technical repository
- Built Random Forest model for ICP classification
- Created enrichment connectors for CSV + LinkedIn data
- Defined segmentation logic (readiness tiers)
- Initial Naive Bayes baseline model
- Basic lead intake + scoring prototype
- Established core architecture patterns
Feedback loop implementation, expanded enrichment sources, model evaluation framework
Ready to Build Smarter GTM?
Join early access or view the GitHub repository for technical details.