Forward-deployed engineering · data onboarding · live
Day one at the client: three systems, three versions of the truth.
The unglamorous 80% of every AI deployment is someone reconciling the CRM, the ERP and an operations spreadsheet that all disagree about who the customers are. This workbench is that fortnight, compressed and governed: every column profiled in seconds, every failing row quarantined by a versioned rule, duplicate entities resolved with the evidence shown and a person on the uncertain calls, and the whole thing closed with an onboarding certificate that states exactly what was done to the client's data. Synthetic client extracts; the mess is planted, the findings are computed. Bring your own CSV to the profiler.
Synthetic client · every transform versioned · humans on the uncertain merges
Pick an extract. Every column is typed, counted and pattern-checked on load: the three date formats hiding in one column, the money that is sometimes "S$1,200.50" and sometimes 1200.5, the nulls dressed up as N/A. This is the conversation with the client's IT team, held with evidence instead of anecdotes.
Deterministic rules, each with an ID and a version, run over every row of every extract. A failing row is quarantined with its reason, never silently dropped and never silently "fixed": the client decides what happens to their data, from a list, not from a shrug.
The same customer lives three times under three spellings. Candidate pairs are scored on name similarity, registration number and email domain, with the arithmetic shown per pair. High-confidence matches merge automatically, with evidence. The uncertain band goes to you: merge or keep separate, and your call is recorded. One pair is a planted trap: two firms that look almost identical and are genuinely different.
Where every surviving field lands: one target model (Customer → Site → Contract), one mapping per source field, one named transform per mapping. The fields that map nowhere are listed too, because "we quietly ignored these columns" is exactly the sentence that detonates in month three.
The onboarding certificate: what came in, what passed, what was quarantined and why, which entities merged on whose authority, which fields mapped under which transform versions. This is the artifact that turns "we cleaned up your data" from a claim into a record.
Runs fully in your browser. Synthetic client extracts from a seeded generator; every finding recomputed on load; uploads never leave this page. · willytai.com