Nexus-Next-Gen EPC Project Intelligence Platform
AI-native platform that unifies vendor verification, schedule risk forecasting, and quality assurance for data centre EPC projects—catching non-conformances, delays and failures before they're expensive.
The Problem
India's data centre pipeline is scaling from ~900 MW (2024) to 2,700+ MW by 2027 — over $15B in capital deployment. Yet 67% of APAC data centre EPC projects overrun schedule by >10% (Turner & Townsend, 2024).
The root cause isn't capital or demand. It's information fragmentation:
- 15,000–40,000 equipment line items and up to 200 concurrent trade contractors per hyperscale project.
- Specs, bids, RFIs, test records, and change orders live in disconnected systems.
- Nobody catches a non-conforming bid clause, a slipping procurement timeline, or a commissioning test failure until it's already expensive.
- Zero error tolerance for Tier III/IV SLA — yet verification is still manual, spreadsheet-driven, and siloed.
The Solution
Nexus is a single intelligence layer with three tightly-coupled AI engines on a shared document/data spine:
Engine | What it does | Key output |
|---|---|---|
Engine 1 — Verification | Checks every vendor bid clause-by-clause against the tender, design standards, and government policy. Numeric requirements (voltage, efficiency, transfer time) are verified algorithmically; complex clauses get LLM reasoning via Claude Opus 5. | Conformance report, deviation flags, vendor risk score, fulfilment ranking |
Engine 2 — Risk Assessment | Five specialised sub-agents monitor procurement lead times, port congestion, monsoon forecasts, workforce availability, and vendor track records. Slip propagates along the critical path. | Risk score per activity, project slip forecast, ranked mitigations with cost/time trade-offs |
Engine 3 — Commissioning QA | Ingests IST procedures (TIA-942, BICSI, Uptime Tier specs), auto-splits checks into machine-verifiable vs. engineer-only, generates test cases, and executes automatable checks via historian telemetry. Failures open NCRs with RAG-sourced precedent resolutions. | As-commissioned quality package, open NCR register |
All three engines communicate via an event bus: an NCR in commissioning instantly lowers the vendor's score in verification; a late shipment holds commissioning test cases; a resolved NCR becomes a precedent for the next project.
Tech Stack & Architecture
Core Stack
Layer | Technology | Notes |
|---|---|---|
Language | Python 3.10+ (standard library only) | Zero external dependencies for the engine code |
Database | SQLite (local) → Aurora Postgres (production) | Same schema, swap-in point at Store.__init__ |
Vector search | TF-IDF cosine similarity | Swap to OpenSearch k-NN / pgvector at Index._vec |
LLM reasoning | Claude Opus 5 on Bedrock | Off by default; deterministic rule checks run without it |
Frontend | Vanilla HTML/CSS/JS | Single-page dashboard + interactive landing page |
Server | Python http.server (local) / Lambda Function URL (prod) | No framework overhead |
Tests | pytest — 67 tests across 7 test files | One file per implementation phase |
System Architecture (The Three Engines)
┌─────────────────────────────────────────────┐ │ COMPANY (EPC Owner / Contractor) │ └───────────────────┬───────────────────────────┘ │ tender issued ▼ ┌─────────────────────────────────────────────┐ │ CLOUD PLATFORM (AWS) │ │ Textract → S3 → Vector/Graph Index │ │ Lambda + Function URL · Bedrock (Claude) │ │ EventBridge · DynamoDB audit · IAM │ └───────────────────┬───────────────────────────┘ │ ┌────────────────────────┼────────────────────────────┐ ▼ ▼ ▼ ┌──────────────────┐ ┌──────────────────────┐ ┌───────────────────────┐ │ Tender Document │ │ Vendor Bids │ │ Govt Policy / Codes │ │ (specs, BOQ, │ │ (technical + comm'l) │ │ (TIA-942, BIS, CEA, │ │ design standards, │ │ │ │ CERC, state DC policy)│ │ client requirements)│ └──────────┬────────────┘ └──────────┬────────────┘ └─────────┬──────────┘ │ │ └────────────┬──────────────┴──────────────┬───────────────┘ ▼ ▼ ┌───────────────────────────────────────────────────────┐ │ ENGINE 1 — SPECIFICATION & QUALITY │ │ COMPLIANCE VERIFICATION ENGINE │ │ RAG-based clause-to-clause matching + policy compliance│ │ + vendor timeline & fulfilment-capacity scoring │ │ Output: Conformance report, flagged deviations, │ │ vendor risk score, audit trail entry │ └───────────────────────┬───────────────────────────────┘ │ approved/flagged vendors, │ PO data, lead times ▼ ┌───────────────────────────────────────────────────────┐ │ ENGINE 2 — PREDICTIVE SCHEDULE & SUPPLY CHAIN │ │ RISK ENGINE (multi-agent) │ │ Inputs: schedule (P6/MSP), procurement status, │ │ shipment tracking, workforce availability, │ │ electricity grid/utility hookup timeline, │ │ global equipment shortage signals, geopolitical & │ │ natural-disaster feeds, commodity/price indices │ │ Output: Critical-path risk score + ranked mitigation │ │ options (not just alerts) │ └───────────────────────┬───────────────────────────────┘ │ site-ready equipment, │ updated schedule state ▼ ┌───────────────────────────────────────────────────────┐ │ ENGINE 3 — COMMISSIONING QUALITY ASSURANCE COPILOT │ │ Ingests test/inspection documents (TIA-942, BICSI, │ │ Uptime Institute Tier specs, IST procedures) │ │ → splits checks: machine-verifiable vs. engineer-only │ │ → auto-generates test cases for both │ │ → executes automatable checks, routes rest to │ │ engineers with generated test scripts │ │ → RAG over cross-site historical failure/resolution │ │ corpus → recommends fixes for failures │ │ Output: As-commissioned quality package + open NCRs │ └───────────────────────┬───────────────────────────────┘ │ ▼ ┌───────────────────────────────────────────────────────┐ │ PROJECT KNOWLEDGE GRAPH + FEEDBACK LOOP │ │ Every finding, resolution, deviation, and delay is │ │ written back — becomes training/RAG data for the │ │ next tender, the next risk model run, the next │ │ commissioning cycle (this project AND future ones) │ └───────────────────────────────────────────────────────┘
AWS Deployment & Runtime Architecture
Lambda Function URL ──► Lambda (python3.12, serve.handler) │ the three engines, one process ┌───────────────┬───────────┼────────────────┬─────────────────┐ ▼ ▼ ▼ ▼ ▼ S3 (versioned) Textract Bedrock DynamoDB EventBridge tender, bids, scanned Claude Opus 5 audit rows, bid.verified, policies, IST pages → clause append-only by risk.assessed, procedures clauses reasoning condition ncr.opened, …Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article