PinkMate is an early-stage research and engineering project exploring deterministic, multi-agent orchestration for B2B go-to-market workflows. It combines classical interpretable machine learning with structured agent-level routing to investigate how lead-scoring, segmentation, and outreach decisions can be made more reproducible, auditable, and transparent than conventional CRM automation or LLM-only agent loops.
Repository scope. This repository documents the technical architecture, agent stack, and reference skeletons for PinkMate's predictive GTM engine. It is a research-engineering artefact — the agent skeletons and orchestration scaffolding are public so the design can be inspected and built upon. Trained model artefacts, production data pipelines, and end-to-end runnable workflows are work in progress and not present in this repository in their current form.
License: Proprietary — source available for evaluation only. See
LICENSEfile for details.
Core thesis. Conventional CRM automation stores leads and applies static rules. LLM-only agent systems generate fluent action sequences but exhibit non-deterministic execution, weak auditability, and unstable explainability. PinkMate explores a middle path: classical interpretable ML for scoring, bounded modular agents for orchestration, and explicit state transitions for reproducibility.
Design choices.
- Multi-agent pipeline: Ingestion → Enrichment → Predictive Scoring → Segmentation → Outreach → Feedback loop, with each agent responsible for a bounded function and observable state transition.
- Predictive ML (planned implementation): classical interpretable models — Naive Bayes, Random Forest, Gradient Boosting — chosen deliberately over opaque deep-learning alternatives to support feature attribution and governance review.
- Agentic orchestration: LangChain / LangGraph-style routing between agents, with explicit state schemas and event-triggered transitions.
- Memory and context (planned): vector-based memory for company, persona, and interaction history, enabling retrieval-augmented context for downstream agents.
- Integration surface (planned): designed to interoperate with Clay, Apollo, PhantomBuster, HubSpot, and Salesforce.
PinkMate is the reference deployment of the Q-MDP orchestration framework named in the paper's final section. The architectural choices here — bounded agent responsibilities, superposition-aware belief state, governance-gated conditional transitions, interpretable classical ML for scoring, and explainability-oriented design — directly instantiate the formal constructs of:
Sehgal, K., & Bhatti, K. N. (2026). Q-MDP: Superposition-Aware State Orchestration for Deterministic Governance of Agentic Systems in Regulated Financial and ESG Infrastructure. Accepted at IEEE Cyber-AI 2026, Bucharest, Romania. Submission #4554.
A full paper-to-code map is in docs/architecture.md. The repository is now in active build against that map on a weekly slice cadence.
pinkmate-agentic-gtm/
│
├─ README.md # This file – high-level overview
├─ requirements.txt # Python dependencies
│
├─ pinkmate/ # Q-MDP formal constructs (Slice 1)
│ ├─ __init__.py
│ ├─ qmdp.py # BeliefState, SuperposedBelief, transition(),
│ │ # governance_commit() — paper §III.A
│ └─ state.py # Pydantic PipelineState, TraceEntry — paper §III.A τ_exec
│
├─ agents/ # Orchestration-graph vertices (paper §III.B)
│ ├─ __init__.py
│ ├─ ingestion_agent.py
│ ├─ enrichment_agent.py
│ ├─ scoring_agent.py # Emits SuperposedBelief; placeholder predictor
│ │ # until Slice 2 wires the real six-model ensemble
│ ├─ segmentation_agent.py
│ └─ outreach_agent.py
│
├─ workflows/ # Orchestrated workflows and routing
│ ├─ pipeline.py # LangGraph StateGraph with governance-gated edge
│ │ # (Slice 1 — authoritative orchestration layer)
│ └─ pipeline.yaml # DEPRECATED; retained for commit-history continuity
│
├─ tests/ # Reproducibility + determinism tests
│ ├─ test_qmdp.py # Unit tests on belief / commit / transition
│ └─ test_pipeline.py # Integration + deterministic-replay smoke tests
│
└─ docs/ # Technical documentation
└─ architecture.md # Paper ↔ code map, slice roadmap
Slice 1 status. The Q-MDP formal layer (pinkmate/qmdp.py, pinkmate/state.py) and the LangGraph orchestration graph (workflows/pipeline.py) are live and executable end-to-end on a mock lead with python -m workflows.pipeline. The ScoringAgent predictor remains a deterministic placeholder in Slice 1 so the Q-MDP orchestration layer can be verified independently; the real six-model ensemble + SMOTE + TreeSHAP wiring lands in Slice 2.
PinkMate's GTM engine is modelled as a set of independent but coordinated agents:
-
IngestionAgent
- Accepts CSV uploads, LinkedIn exports, Apollo data, or direct API feeds.
- Normalises and validates lead records.
-
EnrichmentAgent
- Calls enrichment sources (Clay, Clearbit, Apollo, internal APIs).
- Adds firmographics (industry, size, location, funding) and buyer signals.
-
ScoringAgent
- Designed to apply classical ML models (Naive Bayes, Random Forest, Gradient Boosting) — model-loading logic and training artefacts are pending.
- Will produce a probability of conversion and a confidence score.
-
SegmentationAgent
- Converts scores into tiers (Tier 1 / Tier 2 / Nurture).
- Applies ICP rules (sector, geography, ARR, role seniority, intent).
-
OutreachAgent
- Decides next-best-action: email sequence, LinkedIn DM, nurture cadence.
- Designed to plug into CRMs and outreach tools.
-
FeedbackLoop / Analytics
- Will consume engagement + conversion signals.
- Will adjust thresholds, weights, and features over time.
Orchestration is via a LangChain / LangGraph-style workflow enabling routing, error handling, and retries.
The ScoringAgent class in agents/scoring_agent.py defines the interface contract for lead scoring: a feature dictionary in, a structured score-plus-breakdown response out.
Core responsibilities (planned implementation):
- Load trained classical ML models from
models/. - Accept a dictionary of engineered features.
- Combine model outputs into a final score.
- Return a structured response with
final_score(0–100) and per-modelmodel_scores.
The current implementation contains explicit placeholder return values (e.g. nb_score = 0.78 # simulated) so the interface contract is inspectable before the real model-loading logic is added. This is intentional — the file is a reference skeleton, not a deployed scoring service.
PinkMate is being built against the paper on a weekly slice cadence. Each slice closes a specific gap between the paper's formal claims and the reference deployment.
- Slice 1 (shipped). Q-MDP formal structure wired into LangGraph with deterministic-replay smoke tests.
pinkmate/qmdp.py,pinkmate/state.py,workflows/pipeline.py, test harness, paper-to-code map indocs/architecture.md. - Slice 2. Real ML layer per paper §III.C–D and §V — six-model ensemble (LR / RF / AdaBoost / XGBoost / LightGBM / CatBoost), SMOTE on training partition only, TreeSHAP attribution with σ_SHAP measurement harness, replacing the Slice 1 placeholder predictor in
ScoringAgent. - Slice 3. Durable checkpointing per paper §III.B replay-fidelity claim —
MemorySaver→SqliteSaver→ PostgreSQL checkpointer for cross-process audit-trail persistence. - Slice 4. Empirical validation — latency benchmarks against the
T_fa = O(log n + k)scheduling claim (§III.B), replay-fidelity harness, audit-trace-completeness metric aligned with the ESG companion deployment's measurement. - Later. Integration surface — connectors to Clay, PhantomBuster, Apollo, HubSpot, and Salesforce.
Numbers will be reported only once measured against the slices that produce them.
The architectural and research thesis underlying PinkMate is developed in academic collaboration with Kent Business School, University of Kent. Peer-reviewed work:
- Sehgal, K., & Bhatti, K. N. (2026). Q-MDP: Superposition-Aware State Orchestration for Deterministic Governance of Agentic Systems in Regulated Financial and ESG Infrastructure. Accepted at IEEE Cyber-AI 2026, Bucharest, Romania. Submission #4554 (accepted 18 June 2026).
- Sehgal, K., & Bhatti, K. N. (2026). Comparative Evaluation of Machine Learning Approaches for Minority-Class Financial Distress Prediction Under Class Imbalance Constraints. arXiv:2605.14067 [cs.LG]. https://arxiv.org/abs/2605.14067. Reproducibility artefact on Zenodo: https://doi.org/10.5281/zenodo.20454212 (CC BY 4.0).
- Sehgal, K., & Bhatti, K. N. (2026). Auditable Climate Risk Intelligence from Fragmented ESG Data: Deterministic Orchestration and Imbalance-Aware Learning for Scope 1–3 Validation. arXiv:2606.02604 [cs.LG]. https://arxiv.org/abs/2606.02604. Companion artefact on Zenodo: https://doi.org/10.5281/zenodo.20453252 (CC BY 4.0, OpenAIRE-indexed).
An extended methodological manuscript, A Reproducibility Protocol for Imbalance-Aware Financial Distress Prediction, is currently with Intelligent Computing (AAAS Science Partner Journal with Zhejiang Lab; IF 4.4, Q1) following editorial invitation.
- Product landing page: https://www.pink-mate.co.uk
- Deployment mirror (fallback): https://v0-pink-mate.vercel.app/
- Public concept page (Notion): https://iodized-lily-573.notion.site/Hero-Section-205a5e7da1508078b018e8babfa60c61
Contact / Founder
Karan Sehgal — Founder, PinkMate
AI-native GTM, predictive automation, and growth strategy
karansehgal@pink-mate.co.uk