What we have done.

A tax and accounting software vendor whose users are finance teams at digital-asset funds and quantitative firms.

Every month, users load dozens of spreadsheets exported from exchanges and custodians into the accounting system. Matching headers, identifying the source, splitting trades and reconciling — all by hand.

An agent that works inside the software. It reads the files, finds the header row, identifies the exchange, merges and de-duplicates, maps columns to the standard schema, produces standard transaction rows, then logs in, reconciles and writes. Wherever a judgement is uncertain, the step is flagged and queued for a person to confirm; confirmed judgements become rules, so the next file from the same source is recognised automatically.

Every step is recorded and every judgement carries its evidence. The system is deployed in the client's own environment; model calls go through the channel the client approved; we deliver a deployment package and host nothing.

Running in the client's test environment for the data-cleaning stage. Voucher generation, period close and insight queries are planned next.

How AI becomes acceptable in a domain that cannot tolerate errors: every step verifiable, a person signs off, confirmations become rules.

Founder-led engagement.

A US membership e-commerce company.

The situation many companies face after years of growth. Layered systems — commerce platform, subscriptions and membership, marketing tools, support, behavioural analytics — each with its own copy of the data. An in-house first-generation warehouse had not been designed on dimensional principles; teams pulled from raw tables with inconsistent definitions, and some pipelines ran for hours. The data team knew a rebuild was due.

An overall solution: the enterprise warehouse rebuilt across six domains — merchandising, marketing, operations and supply chain, member services, product analytics, machine learning.

1. Legacy first. Hundreds of hand-written transformations traced back to their sources column by column, with LLM-assisted lineage extraction, and checked table by table.
2. One model. Customer, product, order and date dimensions maintained once and shared by every domain's facts; business keys, hash keys and history on every table.
3. One process per domain. Discovery, reporting requirements, logical model, mapping document, build with test cases, user acceptance, deployment, hypercare.
4. Stack. Snowflake as the warehouse, dbt for modelling, elementary for tests and monitoring, Airflow for orchestration; attribution on unified UTM rules.

Internal client systems; not described here. The point of the case is the method: why a warehouse ends up in that state, what to do first and what next when rebuilding, and what keeps each step correct.

Enterprise-scale data integration under one set of definitions — the large-scale version of the Data Foundation.

The founder's prior work. The founder took part and was responsible for the overall solution.

A Chinese dress brand selling in the US through a Shopify storefront. A few hundred styles in occasion-driven, low-frequency categories — weddings, proms, bridesmaids. Email and SMS through Klaviyo; live chat, loyalty, reviews and try-at-home each a separate app; orders and inventory in a custom ERP. Customer data across six systems; six identities per person.

A stylist persona: on-site pop-ups and live chat, outbound email, SMS and social DMs, a QR code on the package card. Able to recommend, add to cart, trigger checkout and hand off to try-at-home. An operations console for persona, knowledge, campaign scripts and versions. Acceptance on incremental sales.

A shopper views three dresses and leaves. That evening she gets a text, taps through, and the stylist picks up the conversation on site: the occasion, the date, her size, her budget. "A beach wedding in June, not too formal." From a catalogue tagged by occasion, silhouette, neckline, length, fabric, colour and formality, the stylist selects three styles that are in stock, well reviewed and rarely returned, explains why, adds them to the cart, and offers to send them home to try. Checkout happens in Shopify. The post-delivery follow-up and the repurchase prompt two months later come from the same stylist, and every step is recorded against one identity, so she is recognised on any channel.

Three layers, in an order that cannot be reversed.

1. Data foundation. On-site event capture through the site's existing tracking channel; identity resolution across the six systems; a narrow warehouse for orders and products; structured product attributes for a few hundred styles; structured return reasons. Without this layer, "recommend from all customer information" has no input.
2. Segmentation and experiments. One table per person; random, sticky assignment with exposure logging; a hold-out group that is not contacted; cross-channel frequency control. How incremental revenue is computed goes into the contract schedule. The existing twelve marketing-automation flows are twelve active triggers; making them data-driven first produces measurable numbers within two to three months.
3. The stylist itself. One on-site conversation core; one outbound pipeline with sending and compliance left in Klaviyo. Skills are calls to existing functions — add to cart, checkout link, the client's try-at-home flow, discount codes with limits and approval; payment always stays in Shopify. Recommendation is constrained retrieval plus ranking rules: hard filters on what the shopper said, soft ranking on stock, reviews, sales, returns and margin. Product facts come from the systems of record; memory is attached to identity. The console reduces to promotions and FAQ, conversation review and labelling, version rollback, and a dashboard. Twenty to thirty real conversations with "right answers" form the evaluation set, with an upper bound on wrong recommendations.

Phasing: a four-to-five-week paid diagnostic to settle the data inventory, identity rules, tracking, tagging, agent architecture and experiment design; then ten to fourteen weeks for the data foundation and a first-version stylist across pop-up, live chat, email and SMS, with a user-level A/B framework. Social DMs, the package card, virtual try-on and multiple personas in a second phase. The economics were set out in the proposal: annual revenue times a realistic uplift range times gross margin is the client's annual ceiling; the full scope exceeded it, so it was cut down.

Proposal stage. Not signed.

Decomposing "AI shopping assistant" into data, experiments and conversation, in that order; rules over models at a few hundred SKUs; acceptance tied to technical deliverables, with incremental revenue read by the party that earns a share of it. This diagnostic became the prototype of our Data Health Diagnostic.