Four service lines, one standard: you can audit what we leave behind.

Your data estate is years of file shares, vendor drops and databases nobody fully remembers. The usual assessment maps it through weeks of interviews and hand-sampling. We do it in a day, on site, with read-only credentials and nothing installed on your systems.

Lakewright scans the share or bucket from a laptop inside your network, reading the heads of files and never writing to the source. It identifies each feed by evidence (NACHA files, X12 claims, fixed-width mainframe extracts, spreadsheets, scanned PDFs), profiles schemas, flags duplicates and drift, and flags personal data.

The ingestion plans it proposes are plain YAML. Your data owner reviews them and approves the feeds that matter, and the first governed Apache Iceberg tables land the same afternoon. Every finding traces to a file, a digest and a recorded decision.

Take the feeds that matter into production with Lakewright. Claims files, HL7 and FHIR feeds, mainframe extracts, scanned documents and database change streams become governed Apache Iceberg tables inside your own environment, read by Snowflake, Databricks, Microsoft Fabric or anything else that speaks Iceberg.

An AI proposes each ingestion plan and your people approve it. Nothing loads without that approval, and every decision is recorded. Sensitive data is flagged before it reaches the warehouse, and semantic layers are generated so your platform's AI features can answer questions about it.

Lakewright is pre-release. We will tell you plainly whether it reads your formats today, whether that is a week's work, or whether something else would serve you better.

Coding agents make dozens of edits per request, often several requests at once, and git records the diff but not the request behind it. We set your team up on mcpc, version control built for agents, and on the working practices that go with it.

• Every edit snapshotted automatically and linked to the prompt that caused it.
• Undo a whole request, keeping other agents' work.
• Cheap forked workspaces for parallel subagents, merged back three-way.
• Agents, including agents from different vendors, messaging each other about specific functions while the work is still in progress.
• Repository rules that refuse a bad edit before it lands, and project notes in place of one ever-growing instructions file.

mcpc is pre-release and we say so. It has run our own projects in place of git since its first snapshot, imports and exports git so your CI and code host keep working, and ships adapters for Claude Code, Codex CLI, Gemini CLI and Cline. A pilot is still scoped conservatively: one team, one to three repositories that are not business-critical, with a git export kept alongside. The essay lists every known limit.

Most organisations already run more AI than anyone has listed: assistants in products, agents in pipelines, models called from scripts. When something goes wrong, the question is always the same: what did the AI do, at whose request, and can we undo it?

We find every place your organisation calls a model or runs an agent, and with what data, and map each one to the controls you already report against. The gap report says where the record is missing: who or what asked, what was sent, what came back, and what changed as a result. Phase two puts that record in place, so a reviewer can go from an outcome back to the request that caused it.

We apply the same principles our own products are built on, recorded, attributable and reversible, to the systems you already run. This is a review, not an attestation: it does not replace the report an accredited auditor gives you.