Blog
Field notes on data infrastructure, semantic layers and the groundwork AI agents actually need.
- 7 min read
An Ad Creative Engine That Sizes the Test Before It Makes the Assets
An account's weekly conversion volume, not its studio's output, sets how fast an advertiser can learn, and past that ceiling extra variants buy noise rather than knowledge.
Read article → - 8 min read
An Analytics Agent That Cannot Write a Wrong Join, Because It Cannot Write a Join
Prompt instructions cannot stop a model writing a fanned-out join, so this agent's output space contains no SQL at all: it selects governed metrics and the semantic layer writes the query.
Read article → - 7 min read
A Screening Engine That Measures Whether Its Human Oversight Is Real
Nobody builds a machine that rejects candidates; the real hazard is a ranked list nobody reads past, so the number this engine watches hardest is how often a hire comes from below the cut.
Read article → - 7 min read
The Two Catalog Fields No Confidence Score Is Allowed to Fill
Why a marketplace enrichment engine had to split its category schema into attributes a model may infer and attributes only a supplier may declare.
Read article → - 8 min read
A Churn Model That Learned the Wrong Lesson From a Price Rise
Why our best-calibrated risk score was confidently wrong about an entire market, and what changed when we stopped asking one number to describe two unrelated processes.
Read article → - 7 min read
Somebody Else's Name Is Buried in Every Claim File You Close
Why we built the subrogation signal into a claims document intelligence system at intake, and why it is walled off from the coverage view by design.
Read article → - 7 min read
The Note Assistant That Will Not Write a Number the Clinician Did Not Say
Prose survives end-of-day fatigue and digits do not, so this assistant treats measurements as quotations to confirm against audio, never as text to generate.
Read article → - 7 min read
A Price Monitoring Engine That Prices the Move Against the Next Thirty Days
A tactical price match costs a European electronics retailer two things, not one, and the second cost stays invisible unless the engine prices it before anyone approves the move.
Read article → - 8 min read
A Contract Review Copilot That Reads Past the Four Corners of the Lease
Why a commercial property legal team's review copilot had to stop judging clauses on their drafting and start tracking the conditions attached to every right inside them.
Read article → - 7 min read
QA That Scores Every Conversation Without Guessing How Anyone Felt
Scoring every conversation is legally possible only if the rubric never infers a state of mind, which is why the incumbent scorecard had to be rewritten into observable conduct before a single call could be scored.
Read article → - 7 min read
The Forecast Moved, the Order Did Not
Why our first release of explainable replenishment changed nothing, and what happened when we started explaining the number the buyer was actually looking at.
Read article → - 8 min read
The Report Is the Logbook: Voice to Record for Regulated Field Work
A technician's spoken job report becomes an F-Gas logbook entry partway through a sentence, and the capture system has to notice that moment and change its behaviour.
Read article → - 8 min read
A Close Copilot for a Business Where Nothing Is Supposed to Reconcile
Room revenue posts tonight and the cash lands days later, short by a commission, so this close copilot models what should settle instead of hunting for amounts that match.
Read article → - 7 min read
Two Funders Ask the Same Question, and Both Answers Are Correct
Each grant agreement fixes its own indicator definitions, windows and exchange-rate rules, so this engine stores no converted figures and computes every reported number from raw records at render time.
Read article → - 8 min read
An HR Assistant That Cites the Clause and Cannot Tell You Who Asked
Under German co-determination a per-person question log is the regulated object, so the analytics store holds cohorts only and the sensitive routes never invoke a model at all.
Read article → - 8 min read
The Incident Copilot We Built Cannot Touch Production, By Design
DORA's four-hour clock runs from the moment a human classifies an incident as major, so the copilot's real product is a timeline that can prove, months later, when that moment was and what was known at it.
Read article → - 7 min read
An AI Engine That Reads Carrier Invoices and Works Out Who Owes What
A carrier invoice bills one number for fourteen customers' goods, so the engine splits every charge to the shipments that owe it before any match is attempted.
Read article → - 8 min read
An Inbound Agent That Treats Deal Registration as Law
In channel sales an inbound enquiry can already belong to a partner by contract, so this agent arbitrates ownership before anything is scored and refuses to guess when claims overlap.
Read article → - 8 min read
Our Localization Engine Ships Fewer Translations Than It Rejects
Why a meaningful share of interface strings cannot be translated into Polish by anyone at any budget, and what changed when the pipeline started filing bugs against the source code instead of guessing.
Read article → - 8 min read
The Meeting Pipeline That Catches the Scope You Just Gave Away
In professional services the same sentence in the same client call is a pipeline signal and an unbilled-work risk at once, and a pipeline that only writes to CRM fields books the opportunity while losing the margin.
Read article → - 7 min read
A Costing Engine Where the Allergen Matrix Holds a Veto
A cheaper thickener four levels down a prep tree can rewrite thirty-eight allergen declarations, so the allergen matrix vetoes substitutions before ranking ever sees them.
Read article → - 8 min read
A Reporting Engine That Computes Two Correct EBITDAs From One Trial Balance
A lender's EBITDA and a fund's EBITDA are both contractually correct and rarely equal, so every metric this engine publishes names its basis and reconciles to its sibling.
Read article → - 7 min read
A Horizon Scanner That Diffs Provisions, Not Files
A supervisory Q&A row can change a payment firm's obligations without any document changing, so this scanner watches provisions and database entries, not files and feeds.
Read article → - 8 min read
A Tender Engine That Works Out What to Ask Before It Drafts a Word
In regulated procurement nobody may ring the buyer to ask what a requirement means, so the most valuable thing a tender engine produces is the list of questions worth asking while asking is still allowed.
Read article → - 7 min read
A Shared Inbox Where Every Tenant Email Is a Notice
In lettings a tenant's email is a legal instrument the moment it arrives, so the thing worth building is not a triage tool but a register of notices with immutable timestamps.
Read article → - 7 min read
The SOP Assistant That Refuses the Most Common Question on the Floor
The question a floor assistant hears most, whether a dish contains an allergen, is the one it must never answer from a manual, and drawing that boundary was the real design work.
Read article → - 8 min read
A Generator That Splits a 400-Page Manual Into Training Proportionate to the Job
Food-hygiene law requires training commensurate with what each person actually does, which makes splitting a 400-page manual by role a legal act rather than a formatting choice.
Read article → - 8 min read
A Supplier Onboarding Engine Whose Real Output Is an Ownership Graph
Why screening a supplier's name misses the counterparty procurement is obliged to refuse, and why we resolve identity along the chain before adding anything up.
Read article → - 7 min read
A Support Engine That Checks Whether Its Own Draft Is Still True
In subscription commerce the correct reply depends on where the batch cycle stands at the moment of sending, so this engine acts first, writes second, and rechecks its facts at send time.
Read article → - 7 min read
Mining Ten Years of Tickets Without Publishing Workarounds as Answers
Most of an enterprise service desk's resolution history records circumventions, not fixes, and a knowledge base that cannot tell the two apart publishes expired workarounds as answers.
Read article → - 8 min read
An AI Engine That Creates and Optimizes Its Own Content
How we built a self-tuning content pipeline for GoStudent, and why the same approach matters for any brand with a distributed presence.
Read article → - 12 min read
Semantic Layers, Data Warehouses and AI Agents: How We Built GoStudent's Modern Data Stack
How we built GoStudent's modern data stack — and why the same approach matters for any business sitting on years of untapped data.
Read article →