The Challenge
Bancroft Construction manages dozens of active construction projects across multiple regions. Their project data lived in disconnected systems — Procore for field management, spreadsheets for budgets, and manual reports for executive visibility.
Project managers spent hours each week pulling reports, cross-referencing data between systems, and answering ad-hoc questions from leadership. They needed a unified platform that could centralize their data and make it instantly accessible — without requiring everyone to learn SQL or navigate complex dashboards.
The Solution
We built a full-stack construction management platform that integrates directly with Procore's API, centralizes all project data in a structured MySQL database, and layers an AI-powered natural language interface on top. Users simply type questions like "Show me all projects over budget in Manhattan" and get instant, accurate results.
The system learns from every interaction — when users rate queries as helpful, those patterns feed back into the AI's context, continuously improving accuracy over time.
AI Data Explorer
The centerpiece of the platform — a natural language interface that translates plain English questions into validated SQL queries, executes them against the database, and returns formatted results in real time.
Natural Language Queries
Users type questions in plain English — "Show me all active projects in Brooklyn" or "What's the total budget variance this quarter?" The AI understands context and generates accurate SQL.
Security-First Design
Every generated query is validated against a whitelist of allowed tables and columns. Only SELECT statements are permitted — no data modification is ever possible. All queries are logged for audit.
Conversation Memory
The AI maintains conversation context, so users can ask follow-up questions naturally. "Now filter that to only projects over $1M" works exactly as expected.
User Feedback Loop
Every query result includes thumbs up/down rating and optional comments. This feedback directly trains the AI to produce better results over time.
Procore API Integration Pipeline
A fully automated data pipeline that syncs construction project data from Procore's API into a structured MySQL database, keeping the platform's data current without any manual intervention.
30+ API Endpoints
Syncs projects, budgets, budget line items, RFIs, submittals, change orders, change events, commitments, contracts, daily logs, inspections, punch items, meetings, and more.
4x Daily Automated Sync
Scheduled via Windows Task Scheduler at 00:00, 06:00, 12:00, and 18:00 UTC. Lock-file mechanism prevents overlapping runs. Full logging for every sync cycle.
OAuth 2.0 Authentication
Client credentials flow with automatic token refresh. 2-hour token expiry handled transparently. Secure credential management via environment configuration.
Structured Data Storage
All synced data is normalized into a MySQL schema (api_bancroft) with proper indexing, foreign keys, and data type mapping. Supports the AI query layer directly.
Analytics Dashboard
A comprehensive 6-tab analytics dashboard that provides real-time visibility into AI query performance, user activity, error patterns, and the platform's self-learning progress.
Query Performance Metrics
Track success rates, average response times, daily query volume trends, and hourly activity distribution across all users.
User Activity Tracking
See which team members are using the AI explorer most, their individual success rates, feedback patterns, and most common query types.
Error Pattern Analysis
Automatically categorizes and tracks error types, identifies recurring failures, and surfaces patterns that inform AI improvements.
Sync Status Monitor
Real-time visibility into Procore API sync health — last run status, records synced, API call counts, entity breakdowns, and next scheduled sync.
Self-Learning AI Engine
The platform continuously learns from user interactions. Positive feedback reinforces good patterns, negative feedback flags mistakes to avoid, and self-corrections are detected automatically.
Positive Pattern Learning
Queries rated thumbs-up are stored as reference examples. Up to 10 of the best patterns are injected into every AI prompt, teaching it what "good" looks like.
Failure Avoidance
Queries rated thumbs-down or that produced errors are flagged as anti-patterns. The AI is explicitly told to avoid these approaches in future queries.
Self-Correction Detection
When a failed query is followed by a successful one in the same conversation, the system captures both as a correction pair — teaching the AI to skip the mistake.
Graceful Degradation
The learning system is wrapped in error handling so it never breaks normal query processing. If the learning context can't be loaded, queries still work perfectly.
Technology
Built with modern, battle-tested technologies chosen for reliability, performance, and maintainability.
React + Mantine UI
Frontend framework
PHP + MeekroDB
Backend API
MySQL
Database (2 schemas)
Claude AI (Anthropic)
NL-to-SQL engine
Python
Procore sync pipeline
AWS EC2
Cloud infrastructure
Vite
Build tooling
GitLab CI/CD
Version control & deploys
Need something similar?
Whether it's API integration, data pipeline automation, or AI-powered analytics, we can build it for you.