Product & Technology
Leader — Mini-CTO

Problem

Construction teams had BIM models, 2D drawings, schedules, daily logs, RFIs, and project tasks, but they were not connected into a single execution view. IFC/Revit models contained rich element data, but it was locked inside schema-heavy formats. Drawings contained callouts, room tags, title block data, and symbols, but those signals were mostly unstructured. Without a normalized location hierarchy, Procore progress tracking still depended on manual updates and spreadsheet reconciliation.

What you built

• Built a BIM-powered, location-aware construction intelligence platform that turns Revit, IFC, and 2D drawings into structured graph entities.
• Extracted BIM elements such as walls, beams, columns, slabs, doors, equipment, levels, rooms, zones, coordinates, bounding boxes, materials, phases, and custom properties.
• Extracted drawing entities from 2D PDFs using OCR for title blocks and annotations, symbol detection for plan objects, and NLP/NER for sheet metadata and location references.
• Linked BIM elements and drawing references to Procore activities, schedule tasks, RFIs, submittals, daily logs, and progress states.

Results / metrics

• Created a single source of truth for location-aware progress tracking across model elements, drawings, and Procore field activity.
• Enabled queries such as "show unfinished walls on Level 3, Zone B that are blocked by open RFIs" without manual spreadsheet joins.
• Reduced manual mapping effort by using extracted metadata, spatial hierarchy, and fallback zone derivation from grids or bounding box overlap.

Lessons learned

• GUIDs can drift across Revit re-exports, so element identity needs both source IDs and a geometry fingerprint fallback.
• Coordinate systems must be versioned and validated because survey point or project base point changes can silently shift every bounding box.
• Zones are often not modeled explicitly, so the platform needs a confidence-based fallback chain: IfcZone, Revit room, grid cell, then polygon or bounding box overlap.
• The product value is not extraction alone. The value comes from tying BIM, drawings, schedules, and Procore field signals into a decision graph for progress and risk visibility.

Problem

At Procore, project teams were overwhelmed by fragmented data spread across RFIs, drawings, schedules, BIM models, cost systems, inspections, and emails. Users wanted simple answers such as which subcontractor was delaying MEP work or what RFIs were blocking a concrete pour. Traditional dashboards required manual investigation, while early GenAI copilots delivered only about 45% accuracy because they lacked construction context, retrieved the wrong information, and frequently hallucinated. The result was low trust and limited adoption in real project workflows.

What you built

• Feature Agents: introduced domain-specialized agents for schedules, documents, BIM, cost, inspections, and daily logs. Each agent understood its workflow, source systems, permissions, and business logic.
• Custom Agents: launched customer- and role-specific agents for executive portfolio risk summaries, superintendent daily blocker insights, and workflows aligned to customer terminology and KPIs.
• Manifest Agents: built the governance layer defining each agent's capabilities, tools, data access, expected outputs, schemas, and permission boundaries.

Results / metrics

• Improved answer quality from 45% to 90%+ within two quarters.
• Reduced hallucinations by more than 60% and doubled retrieval precision.
• Improved response times by 35% through smarter routing and agent specialization.
• Reduced token costs by 28% through routing optimization, prompt compression, and caching.
• Moved AI from pilot experiments into daily construction execution.

Lessons learned

• Strong data matters more than clever prompting.
• Specialized agents outperform a single large assistant for complex construction workflows.
• Governance through manifests is essential for enterprise trust, security, and scale.
• Evaluations and observability turn AI from a demo into a measurable, continuously improving product.

Problem

Enterprise SaaS platforms at Autodesk and Procore needed to support high-volume transactional and operational workloads across thousands of customers, regions, and downstream systems. At Autodesk, the challenge was subscription commerce: orders, pricing, entitlements, renewals, partner channels, SAP, and Salesforce synchronization. At Procore, the challenge was operational intelligence: project health, regional rollups, financial risk, RFIs, daily logs, schedules, and nested 360-degree company and project profiles. Both problems shared the same platform question: how do you build a multi-tenant SaaS architecture that scales predictably, keeps data fresh, supports downstream use cases, and remains performant under peak load?

What you built

• Designed transactional subscription APIs for Autodesk.com and partner channels to create orders, validate pricing, manage products, and trigger subscription entitlements at scale.
• Built event-driven integration flows connecting customer and partner actions to SAP order management, Salesforce subscription records, entitlement systems, and downstream analytics.
• Built a 360-degree project and company profile platform at Procore that aggregated operational metrics across RFIs, Daily Logs, Financials, PMQS, and Preconstruction.
• Evolved the data architecture from DynamoDB key-value metrics to DocumentDB and MongoDB-style document aggregation for nested JSON workloads requiring low-latency rollups.

Results / metrics

• Supported Autodesk's subscription platform transformation contributing to approximately $1B ARR.
• Enabled scalable partner order placement, subscription lifecycle changes, and entitlement synchronization across commerce, ERP, and CRM systems.
• Supported executive dashboards with a target latency under 250ms for project, company, country, and region-level rollups.
• Improved the foundation for downstream analytics, ML, alerts, and customer-facing operational insights.

Lessons learned

• A database choice is a product decision when latency and customer trust are part of the value proposition.
• DynamoDB works well for simple key-value access, but rich runtime aggregations over nested operational profiles need stronger document-query patterns.
• Multi-tenancy requires product-level policies for isolation, throttling, fairness, observability, and geo-aware compliance.
• Event-driven architecture is essential when every operational change must power downstream analytics, alerts, ML, and customer workflows.

Problem

Enterprise customers with over $500M ARR were threatening non-renewal because cross-domain reporting did not exist. Financials, PMQS, and Resource Management data were fragmented, and full materialization required approximately 45 minutes.

What you built

• Built a cross-domain executive reporting layer across major Procore product lines.
• Introduced customer-aware partitioning for the top enterprise accounts.
• Migrated data access patterns from DynamoDB to DocumentDB to native MongoDB where needed.

Results / metrics

• Reduced executive reporting freshness from approximately 45 minutes to 5 minutes.
• Achieved under 250ms P99 query latency for critical reporting paths.
• Protected high-value renewals by giving customers reliable portfolio visibility.

Lessons learned

For executive workflows, perfect freshness is not always the right goal. The better product decision was a clear consistency tradeoff: bounded staleness, predictable latency, and trustworthy metrics.