AI Lead Conversion Platform

English | 日本語

A privacy-safe reconstruction of a lead-conversion prototype developed during a 2024 AI hackathon in Ulm, Germany.

The current 2026 reconstruction now includes the main product slices needed to review a synthetic lead, run a reconstructed historical prediction flow, and request a human-reviewed outreach draft without publishing the original corporate dataset or proprietary source code.

Important

The original 2024 prototype was a team project and received a hackathon award. This repository is an independent 2026 reconstruction. The original corporate dataset, internal documents, credentials, and proprietary source files are not included. The reconstructed repository itself is not the award-winning artifact.

Why this project exists

Sales teams often have more leads than they can review manually. This project explores how a team could prioritize leads, understand the evidence behind the reconstructed model flow, and draft a reviewable outreach message without exposing personal or corporate data.

Current product flow

Synthetic CRM data
  -> Historical preprocessing reconstruction
  -> Recovered 18-feature modeling table
  -> Reconstructed model training / prediction
  -> FastAPI historical prediction boundary
  -> React / TypeScript review dashboard
  -> Human-reviewed outreach draft API

The dashboard currently displays a synthetic lead and a fixture-backed historical prediction label, then loads an outreach draft from POST /historical/outreach-draft.

The frontend does not yet call POST /historical/predict directly, load live lead data, send messages, configure an external LLM provider, or reproduce the exact 2024 visual design.

Implemented reconstruction slices

• Public/private data boundary documented
• Historical lead/note raw shape and join relationship reconstructed
• Public aggregate CRM profile documented
• Privacy-safe synthetic generator calibrated to observed aggregate behavior
• Historical preprocessing stages reconstructed with explicit evidence boundaries
• Recovered 18-feature modeling table assembled
• Deterministic stratified train/test split
• Prior baseline model primitive
• Historical XGBoost training and label prediction primitives
• Historical XGBoost accuracy calculation and feature-importance extraction
• Historical Random Forest training primitive
• POST /historical/predict FastAPI endpoint
• React / TypeScript / Vite historical review dashboard
• Privacy-safe outreach prompt contract
• Deterministic local LLM-adapter fallback
• POST /historical/outreach-draft FastAPI endpoint
• Dashboard connection to the outreach-draft endpoint
• Connect the dashboard prediction display to POST /historical/predict
• Add a consolidated reproducible experiment report
• Add broader model-comparison metrics and diagnostics
• Add Docker Compose / CI workflow
• Optional external LLM-provider integration behind the existing adapter boundary

See the roadmap for planned milestones and the repository's historical-reconstruction boundaries.

Reconstruction principles

• Ground the public schema and aggregate distributions in evidence visible in public 2024 hackathon artifacts without publishing source records.
• Preserve the distinction between wide raw CRM data and the smaller modeling dataset rather than designing directly around today's feature subset.
• Keep inferred historical behavior clearly labeled as inferred rather than verified.
• Prevent entity and target leakage by splitting at lead level and excluding post-outcome information.
• Keep preprocessing fitted on training data separate from evaluation data.
• Treat generated customer messages as human-reviewed drafts, not autonomous decisions.
• Prefer reproducible tests and explicit evidence boundaries over unsupported historical claims.

Quick start

python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[dev]"
uvicorn lead_intelligence.api:app --reload

Then open http://127.0.0.1:8000/health.

Run the Python tests with:

pytest

Run the dashboard separately:

cd frontend
npm install
npm run dev

During local development, Vite proxies /historical/* requests to the FastAPI service on http://127.0.0.1:8000.

Synthetic CRM data

The original public notebook workflow shows a wide leads.csv export (86,244 rows x 181 columns) and a one-to-many lead_notes.csv export (134,793 rows x 16 columns). The two were joined through ObjectID = ParentObjectID, then reduced to a smaller working dataset.

The reconstruction keeps that separation. It also calibrates several synthetic properties from public aggregate notebook outputs, including:

• the five observed workflow status proportions;
• all 39 observed Source_Text frequencies;
• the share of leads with note rows and average note count;
• field-specific missingness for the historical working cohort;
• observed cardinality/skew for owner, sales-unit, territory, and lead-name fields;
• mixed date formatting, language proportions, and dirty note placeholders.

All public records are newly generated. Historical identifiers, customer or employee names, and free-text notes are not copied, masked, translated, or sampled.

The reconstruction is therefore aggregate-calibrated synthetic data, not an anonymized copy of the original corporate dataset. Properties whose full public distribution is unavailable remain explicit approximations rather than inferred historical facts.

Small matching examples are committed at:

• data/synthetic/leads_sample.csv
• data/synthetic/lead_notes_sample.csv

See Synthetic CRM specification and Historical CRM aggregate profile for the source evidence, calibration boundary, and remaining approximations.

Repository boundaries

• src/lead_intelligence/: Python application, reconstruction, and ML code
• tests/: automated tests
• frontend/: React / TypeScript review dashboard
• data/synthetic/: generated, non-identifying samples only
• docs/: project history, data policy, roadmap, and engineering decisions

For the historical boundary and attribution policy, see Original project context. For handling rules, see the Data and secrets policy.

Tech stack

Python · FastAPI · pandas · scikit-learn · XGBoost · React · TypeScript · Vite

License

No license has been selected yet. Copyright, licensing, and attribution boundaries will be reviewed before any external reuse or distribution.