Designing StoneAge: Multi-Model Routing (GPT-5, GPT-4o, Gemini) & Why Search-Plus-Rules Beats Fine-Tuning in Systems of Record
Published November 16, 2026 · Lead AI Architect (Dubai, UAE)
1. Meet Users Inside the System of Record
Standalone internal chat portals suffer steep adoption drop-off because engineers must copy-paste data out of their authoring tools, lose document hierarchy, and manually re-verify compliance rules. When designing StoneAge at John Deere, we embedded the AI writing copilot directly inside the company's engineering system of record using a React frontend, FastAPI backend, Okta OAuth 2.0 authentication, and stateful LangGraph agents.
2. Why We Chose Search-Plus-Rules Over Weight Fine-Tuning
Stakeholders frequently ask whether to fine-tune an LLM on internal service manuals. In regulated manufacturing, pharma, and enterprise ERP environments, encoding factual product knowledge into model weights creates an auditability liability:
- Instant Knowledge Revocation: When an engineering bulletin updates a torque specification or safety procedure, updating an OpenSearch hybrid index and deterministic rule table takes minutes — whereas weight fine-tuning leaves stale facts baked into parameters.
- Line-Item Attribution: Hybrid lexical + semantic retrieval paired with rule checks guarantees every drafted section traces back to approved source content and flags safety-critical wording automatically.
- Model Agility via Task-Level Routing: Because domain knowledge lives cleanly in retrieval and rule layers rather than locked inside one fine-tuned checkpoint, StoneAge dynamically selects the best foundation model per subtask across GPT-5, GPT-4o, and Gemini — optimizing accuracy, multimodal reasoning, latency, and FinOps cost simultaneously.
3. Zero-Trust Write Boundaries
Reading enterprise context and mutating official records require fundamentally different security postures. In StoneAge, agents can propose drafts, highlight safety-critical phrasing, and validate standards autonomously — but can only modify official engineering records after a fresh, explicit user approval, with every attempt written to an immutable audit log. For open-source implementations of transactional agent boundaries and compensating ERP rollbacks, see aegis-runtime.
About the Author — Nagarjuna Reddy Kamireddy
Nagarjuna Reddy Kamireddy (Nagarjuna Kamireddy / Nagarjuna K) is a Lead AI Architect, AI Solutions Architect, and Corporate AI Trainer in Dubai, UAE. Connect on LinkedIn or email gmailtonagarjuna@gmail.com.