Not what was built - what it changed.

The reports already held everything - sales volumes, pricing trends, deal flow, year-on-year movement. What was missing was the sentence a busy reader actually wants: what does this mean, right now, for this borough and this quarter. Power BI can do that natively, but only through Copilot on costly Fabric capacity, and the output is generic at that.

So I built it directly into the report instead. A layer of logic gathers the KPIs that matter, sends them to an AI model with precise editorial instructions, and renders a polished narrative back into the page. Change a filter and the summary rewrites itself in seconds - for a market analyst, an executive, or a broker, depending on how it is briefed.

It runs inside the existing environment with no extra Microsoft licensing, and because it is driven by ordinary API usage, the running cost is a rounding error next to the licence it replaces - while reading as if written by someone who genuinely knows the market.

A reporting database of several million rows needed to move from one managed cloud platform to another. A migration like this is rarely just a copy: the data has to land intact, the model has to keep performing, and every downstream report has to keep telling the truth.

The subtle part was time. Timestamps were being converted differently depending on where the report ran, so the same figure could differ between the desktop and the published service. I pushed a daylight-saving-aware conversion down into the database itself, so every environment returns an identical, unambiguous value - closing a discrepancy that is easy to miss and hard to trust once found.

With the gateway, encryption, and connection details set up properly, the result was a migration that was not just complete but dependable - the same numbers everywhere, every time.

Power BI ships with a natural-language Q&A feature, but it answers with charts and tables, and it guesses at what you meant. For someone who just wants the average cap rate for a neighbourhood, a bar chart is friction, not an answer.

I built a compact custom visual instead. It offers a curated set of the questions people actually ask, computes each answer directly from the filtered data, and returns it as clean, formatted text - a percentage with its sample size, a ranked list with property names and values. Every answer is deterministic: the same data and filters always give the same result.

It computes everything locally, so there is no external service, no latency, and no per-query cost. It reacts to slicers like any native visual, and because it is a proper TypeScript custom visual it runs everywhere Power BI does, including Publish to Web and Salesforce. Under the hood it pages through the full dataset and de-duplicates at the row level, so aggregates are always based on the complete picture.