Sofvie Inc.

Sofvie Inc. is a North-America-headquartered SaaS provider focused on industrial and mining safety and operational analytics, serving industrial-safety customers across multiple geographies. It runs a multi-tenant platform on a database-per-client isolation model. The agentic workload in this case study is embedded directly inside Sofvie's commercial product and exposed to in-product analytical users at each client tenant.

Sofvie's analytical users needed to ask questions of their data in plain language, but that data spans roughly 40 tables behind 800+ existing Django REST API endpoints, under strict per-tenant isolation and privacy-relevant PII columns. A single-shot LLM prompt was infeasible: it could not reason reliably over the multi-table schema, it ballooned cost by sending the full schema on every prompt, and it offered no deterministic validation chokepoint. Sofvie's starting point was a proof-of-concept that proved a tool-using agent loop was viable but lacked every production control — no authentication, no tenant isolation, a regex-based SQL check, and a single internet-exposed host.

VeUP designed an MCP-orchestrated agent on Amazon Bedrock rather than a long-context prompt or a single-shot model. A user question enters an async agent runtime; a first Bedrock invocation produces a tool plan, and the agent then issues iterative tool calls to a bespoke Model Context Protocol (MCP) server that wraps Sofvie's 800+ existing Django REST endpoints. The team deliberately chose to wrap those endpoints rather than have the model generate SQL directly — a design pivot that kept the model's job to intent-to-tool matching and made cost-optimized Amazon Nova Lite economically viable. A second Bedrock invocation synthesizes the redacted result into a natural-language answer. The measured production cost signature is roughly 2 Bedrock invocations plus ~42 MCP tool calls per question — the telemetry of a genuine agentic loop.

Model selection was evaluated for tool-using agent workloads: Amazon Nova Lite as the cost-optimized primary, Amazon Nova Pro as the higher-reasoning fallback, and Anthropic Claude on Bedrock retained as a qualified A/B and disaster-recovery alternate over the same MCP tool surface. Keeping inference on Bedrock gave Sofvie a single AWS-native control plane for model-invocation logging, IAM, and KMS — with no GPU capacity planning for a team with no ML-platform function.

Security and Responsible-AI controls are structural, not bolt-ons. Authorization is identity-scoped tool authority: the agent never holds a database handle — the runtime resolves a signed JWT to exactly one tenant database and binds every tool call to that read-only, SSL-required connection, so cross-tenant access is structurally impossible rather than policy-enforced. The MCP server is a single safety chokepoint that AST-validates every proposed SQL string against an allow-list and redacts PII columns before any result re-enters the model's context. A prompt-injection adversarial corpus and a 100+ representative-question golden regression set run in CI on every prompt, model, or tool-surface change, with a thumbs-feedback human-in-the-loop signal. The production agent went live on 2026-05-13, 28 days ahead of the contracted close, with zero cross-tenant incidents at go-live.