Case Study 01 · John Deere (2023–2026)
StoneAge: Multi-Model Agentic Writing Copilot Inside the Engineering System of Record
Part of 750K+ FY26 Automated Requests
Multi-Model Routing (GPT-5, GPT-4o, Gemini)
Safety-Critical Wording Guardrails
Explicit Human-Approved Record Writes
Problem: Technical writers authoring and maintaining complex heavy-machinery service manuals needed AI assistance directly inside the company's engineering system of record (PLM environment), where hallucinated torque specs, unapproved terminology, or missing safety warnings are unacceptable.
Architecture & Solution: Architected StoneAge as a multi-agent LangGraph copilot integrated with React, FastAPI, OpenSearch hybrid lexical+vector retrieval, and Okta SSO:
- Task-Specific Multi-Model Router: Dynamically routed drafting, structural synthesis, and verification steps across GPT-5, GPT-4o, and Gemini based on reasoning complexity, context window, and latency budget.
- Search-Plus-Rules Over Fine-Tuning: Deliberately avoided model weight fine-tuning so rapidly evolving product specifications remain 100% auditable and immediately updatable via approved OpenSearch indices and deterministic rule engines.
- Safety & Auditability: Flagged safety-critical wording automatically; enforced a hard architectural invariant where agents can mutate official engineering records only after a fresh, non-cached human approval, logging every attempt.
Case Study 02 · John Deere · GenAI Rescue
Rescuing a Failing ~30%-Accurate GenAI Product to ~70% Production Accuracy
~30% → ~70% Verified Accuracy (+40 pts)
46 Daily Production Users
GPT-4o Vision Diagram Parsing
Problem: An existing single-prompt, single-model GenAI application failed on dense engineering schematics and product-family variations, stalling at roughly ~30% accuracy and losing user trust.
Architecture & Solution: Decomposed the monolithic call into a deterministic step-by-step agentic workflow:
- Stage 1 — Multimodal Diagram Ingestion: Extracted callouts, part topology, and spatial relationships from engineering diagrams using GPT-4o Vision.
- Stage 2 — Product-Family Routing: Classified equipment family and routed context strictly to family-specific domain rules and terminology stores, eliminating cross-model contamination.
- Stage 3 — Rule-Engine Verification Gate: Checked every structured output against deterministic domain constraints before surfacing to authors — lifting production accuracy to ~70% and driving active adoption across 46 specialists.
Case Study 03 · John Deere · Evaluation, GraphRAG & Scale Automation
AIFlow (255 Quality Gates), 6-Agent Neo4j GraphRAG & High-Volume Support Automation
255 Automated Eval Checks
~171K Broken Links Repaired (~8,600 hrs)
~110K Compliant Docs (100% Large Tractor Adoption)
~25% Support Tickets Automated (~83% Accuracy · ~4,700 hrs)
Connected Knowledge (6-Agent Neo4j GraphRAG): Parsed hierarchical engineering manuals without flattening structural hierarchy and built a 6-agent authoring team backed by a Neo4j knowledge graph so every synthesized technical response carries traceable paragraph- and node-level citations.
Measurable Quality (AIFlow): Built AIFlow, an internal library of 255 quality, faithfulness, safety, and formatting checks integrated directly into CI/CD pipelines to block deployments whenever answer quality regresses — paired with asynchronous production telemetry tracking accuracy, drift, cost, and latency with zero user-facing latency penalty.
High-Volume Production Automation:
- Automatically repaired ~171,000 broken cross-document engineering links (~8,600 hours saved — equivalent to ~3 FTEs per year).
- Generated ~110,000 standards-compliant documents, achieving 100% adoption across the Large Tractor business unit.
- Automated ~25% of incoming support tickets end-to-end at ~83% accuracy, routing uncertain confidence bands cleanly to human specialists (~4,700 hours saved).
Case Study 04 · Cloud Platform & Responsible AI Governance
Self-Service AWS AI Deployment Platform & Enterprise AI Governance Council
Adopted Department-Wide + 3 Partner Orgs
12+ Engineers Led · 300+ Staff Enabled
NIST AI RMF & EU AI Act Alignment
Built a standardized, self-service AWS AI deployment platform using Terraform and GitHub Actions (ECS Fargate, Lambda, S3, DynamoDB, Athena, IAM) adopted across the home department and three partner organizations while owning AWS FinOps cost governance. Served on the enterprise AI Governance Council applying NIST AI RMF and EU AI Act risk tiers to decide which AI initiatives to scale, remediate, or retire.
Case Study 05 · Persistent Systems, GSK (UK) & PETRONAS / Deloitte
From 48-Hour Batch Loads to 12-Minute SAP HANA Pushdown & Enterprise MLOps
PETRONAS Load: 48 Hours → 12 Minutes
GSK UK: In-DB HANA ML (PAL / APL / R)
6 Production Client ML Systems at Persistent
- Persistent Systems (Engineering Lead — ML, 2021–2023): Delivered static/dynamic model serving, optimization, and secure MLOps pipelines across PIAM physical identity access recommendation, iDataCleanse entity deduplication, computer-vision face recognition, VMS recommendation, VectorFlow POD, and Vertina NLP.
- GlaxoSmithKline (GSK), Brentford, UK (HANA Lead Consultant, 2014–2016): Pioneered code-to-data pushdown by running regression, random forest, SVM, clustering, ARIMA/Holt-Winters forecasting, and churn scoring directly inside SAP HANA (PAL, APL, embedded R), shrinking MRP and financial consolidation batch cycles from hours to minutes.
- PETRONAS Energy Trading & Deloitte, Kuala Lumpur (HANA / BPC Architect, 2011–2012): Re-architected an energy-trading data pipeline in SAP HANA, cutting load execution from 48 hours down to 12 minutes (240× speedup) and securing full internal rollout approval.