Lead Data & AI Engineer · Azure · .NET · Microsoft Fabric

20 years building enterprise production systems on the Microsoft stack — 15 of them on Azure — for regulated environments where compliance is not optional. Now working at the boundary where AI features have to reach production safely: deterministic guardrails, auditability, and honest scoping.

📍 Based in Germany · permanent residence (Niederlassungserlaubnis) — unrestricted employment and self-employment, no sponsorship required 🌐 English-speaking teams · remote-first, open to hybrid

I publish reference architectures with the design reasoning and the evaluation kept in the open — not just working code, but the decisions behind it and the evidence that it works. Both repos label what is built versus what production would additionally require.

.NET 8 · Azure Functions · Azure OpenAI · React — a serverless pipeline turning business data into executive insight, on a deterministic-before-probabilistic design: statistics and anomalies are computed in C#, and the model narrates a bounded, pre-computed summary. Raw rows never reach the LLM.

• Secretless end to end — user-assigned managed identity for OpenAI, Blob and SQL; user-delegation SAS for uploads, so no account key exists anywhere in the path
• Prompt-injection mitigation — user-derived strings sanitised and fenced as untrusted data, with a fence-breakout regression test
• Append-only audit trail — every attempt persists its signals, the exact summary the model saw, the raw response, the prompt version and token counts
• Production-shaped reliability — atomic job-state transitions, idempotent triggers, lease-based recovery of crash-orphaned jobs, retry caps, adaptive token budgeting
• Tests gate deployment — xUnit over the deterministic core; GitHub Actions runs them on every push and deploys only if they pass, via OIDC federated identity with no deployment secrets

Microsoft Fabric · Delta Lake · DirectLake · Azure AI Foundry — a medallion reference architecture for private-equity / venture-capital analytics, built and validated end to end on a live Fabric tenant.

• Bitemporal conformed layer — effective_date versus ingestion_date, Type-2 SCD for restatements, so "as-of" queries are first-class
• Reconciliation that surfaces conflicts rather than silently picking winners — scored 1.000/1.000 against a synthetic conflict oracle
• Gold star schema → DirectLake semantic model → Power BI, with a published measure contract covering definitions, valid grain and caveats
• AI layer — route-then-invoke fusion agent over an Azure AI Foundry vector store and function-calling retrieval against the semantic model, with an oracle-based evaluation harness rather than LLM-as-judge (structured leg 6/6 grounded; document leg 4/6, with citation-accuracy and annotation coverage reported separately)
• 17 numbered design decisions (DD-01 → DD-17), including decisions that were later revised, with the revision reasoning kept

Portfolio-scale build on a trial tenant against a synthetic corpus — not client work, and labelled as such throughout.

Two decades across regulated financial services and enterprise clients — private equity and investment management, alternative investments, global custody, energy-trading compliance and MAR trade surveillance, SWIFT payment integration, and connected-vehicle telemetry. SQL and relational modelling since 2006; 15 years on Azure across event-driven, serverless and microservices systems.

Certifications: DP-700 (Fabric Data Engineer Associate) · MCSE: Cloud Platform & Infrastructure · MCSD: Azure Solutions Architect

Contract or senior permanent, remote-first — Azure and .NET platform engineering, data platform, and applied AI. Regulated financial services is where I have the deepest domain context, but the stack matters more than the sector.

LinkedIn · vinodrpatil@outlook.com