• 20 years in tech [ C/C++ → Java/Scala/C# → Python → Agent ( context/intent/spec ) ] — roughly 4 dog years, currently in production
• Currently evolving from legacy specialist to AI generalist — making AI understand what you meant, and making legacy systems understand it's time
• Working with Tier-1 European banks/ regulated enterprises on the three problems that quietly block enterprise AI
  • Context you can't Trust.
  • Estates too brittle to Modernize.
  • Agents that Drain Wallets and Jump Rails.

📍 London, UK · LinkedIn

• Context & Intent Engineering
  Designing the context supply chain enterprise AI actually runs on — semantic models, data contracts, and retrieval pipelines that give agents governed, auditable answers instead of confident guesses. Garbage in, hallucination out. This is the fix.

• Agent Governance & Cost Control
  Architecting the governance and cost controls that keep agents inside their budget and inside their brief — policy enforcement, loop guardrails, model routing, and intent-encoded stop rules. Most agent cost and compliance failures share a root: no one specified where the agent stops.

• Analytics Estate Exit (SAS · Alteryx)
  Migrating SAS workloads — Base SAS, SAS/STAT, DI Studio, Model Manager, Visual Analytics, Viya, and AML suites — to Databricks and Snowflake via automated transpilation and semantic uplift. Statistical logic preserved; platform lock-in eliminated.

• EDW Exit (Teradata · Netezza · Oracle · SAP BW)
  Large-scale EDW migration to Databricks and Snowflake — schema translation, workload profiling, query optimization, and medallion architecture design. The on-premise performance tax: cancelled.

• ETL & Integration Modernization (DataStage · Informatica · ODI · Talend)
  Retiring legacy ETL and re-engineering pipelines as cloud-native data products — embedded data quality, machine-readable lineage contracts, observable orchestration. Opaque batch jobs not invited.

• Mainframe & Batch Orchestration Exit (COBOL · JCL · AutoSys · Control-M · OPC · UC4)
  Extracting decades of embedded business logic from mainframe-bound workloads into testable, versioned, agent-consumable services — SLA contracts preserved, full observability gained.

• Static BI → Conversational Analytics or Vibe Analytics (Tableau · Cognos · SAP BO · MicroStrategy)
  Migrating fixed dashboard estates to governed conversational analytics — Databricks AI/BI Genie, Snowflake Cortex Analyst, Microsoft Fabric, and MCP-based agents (Copilot, Claude) — with semantic models as the governed query surface and natural language as the interface. The three-day dashboard request that answers the wrong question: cancelled.

• Regulatory Reporting & Data Lineage (RegTech)
  BCBS 239, Basel III/IV, FRTB, CRR3, COREP — modernized with machine-readable lineage contracts that satisfy Principle 2 automatically and produce data assets an agent can reason over without apology. Regulators don't accept "it was like that when we got here".

• Vendor lock-in has moved to the semantic layer
  Designing exit-readiness into vendor-managed semantic layers before migration commits, not after. Snowflake Cortex, Databricks Unity Catalog, and Google Agentic Data Cloud all carry the same lock-in surface as the platforms they replaced. New labels, familiar trap.

• The universal semantic layer isn't coming
  Coordinating semantically-driven artifacts across a heterogeneous estate instead — the universal semantic layer has been eighteen months away for a decade. OSI compliance is the line every target architecture has to cross from 2026 onwards.

• Semantic drift is a compliance event, not a data quality one
  Detecting drift in agent-interpreted business logic before it surfaces as a Principle 2 breach, not after the auditor asks why the number changed. Auditors don't accept "the model seemed confident".

• Conversational BI breaks on real estates, not demo data
  Most NL-to-insight demos work on a clean star schema with five tables. Enterprise estates have 50,000 tables, three conflicting definitions of "revenue," and a semantic layer nobody has maintained since 2019. Genie, Cortex Analyst, and Fabric Copilot are only safe in front of an executive when the semantic model underneath is governed. Almost none are.

• Most cloud migrations just relocate the problem
  Agent-ready target architectures where governed semantic models are the query surface, not raw tables. One migration enables AI. The other moves the problem to the cloud at three times the cost.

• Nobody designed where the agent stops
  Where governance and identity belong in the agent stack — AgentHarness (LangGraph, CrewAI, AG2) runs agents; AgentOS (Cedar policy, MS AGT, Unity AI Gateway, AgentCore) governs them. Conflating the two is how compliance debt gets designed in from day one. Ignoring both is vendor lock-in with better branding.

Databricks Snowflake Microsoft Fabric AWS Python SQL Knowledge Graphs LangGraph MCP Cedar MS AGT