Four overlapping domains. Most engagements touch more than one.

The substrate. Data preparation for AI in operational environments: canonical data models, entity resolution, taxonomies, ontologies, and the cleansing and normalization that comes before any of it can be used. This is where most operational AI projects quietly fail, not because the model is wrong, but because the representation of the business is incoherent. Two suppliers under three names. Three classification schemes for the same product family. Six sources of truth, none of them aligned.

What this looks like in an engagement

• ·Canonical models for items, suppliers, customers, contracts, accounts
• ·Data preparation for AI use cases: master data cleansing, deduplication, schema reconciliation, and the entity resolution that turns six sources of truth into one
• ·Entity resolution pipelines: deterministic, probabilistic, or hybrid
• ·Taxonomies and ontologies designed to be used by production systems, not just documented
• ·Normalization of fragmented operational data into structures that retrieval and agents can reason over

The substrate is only useful if the systems above it can find what's there. Vector search alone underperforms in operational domains; structured retrieval over canonical models (graph traversal, hybrid retrieval, semantic search with type constraints) is what makes operational AI feel intelligent rather than approximate.

What this looks like in an engagement

• ·GraphRAG and hybrid retrieval architectures
• ·Internal and semantic search systems
• ·Knowledge graph design, implementation, and maintenance
• ·MCP servers that expose operational systems to agents in a governed way

Agents are easy to demo and hard to deploy in environments with real consequences. The interesting questions are governance ones: where does human approval belong, how do you audit decisions, what happens when the agent is confidently wrong, and how do you keep the loop tight enough to actually learn from production.

What this looks like in an engagement

• ·Governed agent architectures with explicit human-in-the-loop checkpoints
• ·Operational copilots embedded in real workflows, not standalone chatbots
• ·Tool-use design and approval pattern engineering
• ·Evaluation loops and feedback systems that compound over time

The unglamorous infrastructure that turns architecture into production. Most engagements need it, so I do it, not as a separate service line but as the connective tissue between design and operation.

What this looks like in an engagement

• ·Ingestion, transformation, and serving on Databricks, Snowflake, or your existing data platform
• ·Evaluation harnesses with regression detection
• ·Observability for retrieval quality, agent behavior, and data freshness
• ·Rollout patterns: shadow mode, partial rollout, kill switches