Pulse answers the question every dashboard leaves open: revenue dropped, but why? It reads a company's orders, customers, payments and support tickets, spots what changed, and explains it in plain language, with every number coming from the database rather than from the AI.
Most businesses have the data but not the answers. Sales sit in one system, payments in another and support in a third, with duplicates, mixed date formats and five spellings of the same country.
A normal dashboard shows that revenue fell. It does not say which category, which channel or which customers caused it, so someone spends a day in spreadsheets to find out.
Adding a chatbot on top usually makes it worse. A language model will happily invent a number that sounds right, and a business cannot act on that.
About 30 SQL functions compute every KPI, trend, funnel and retention cohort directly in Postgres, under row level security. Revenue is defined once, so every breakdown adds up to the headline number.
13 generic detectors look for category declines, refund spikes, payment failures, inactive VIPs and more. They are not told about the ten stories planted in the demo data. They find them.
The AI can only call 17 whitelisted analytics tools with validated arguments. The cards, tables and charts in its answers are built from those tool results, never from the model's text.
A Data Quality Center that scans 17,000 raw records from four sources and cleans them, a CSV importer that maps and validates columns, and an Action Center fed by six automation rules that run inside Postgres.
Making the AI trustworthy. The model never writes SQL and never supplies a number. It chooses tools, each tool runs a fixed, parameterised SQL function under the visitor's own permissions, and the figures on screen come straight from those results. If a model cannot call tools, a keyword planner fetches the data instead, and if the provider fails, the page still shows the data it already has. The other hard part was the public demo: every visitor gets an anonymous, isolated session, so they can create actions, notes and reports without seeing anyone else's, and Reset wipes only what they made.
Live as a public demo at pulse.ninga.dev. Anyone can open a private session on Northstar Commerce, a fictional outdoor retailer with about 25,600 customers and ten business stories hidden in its data, and ask the assistant what is going on. The same schema already supports real organisations, roles and row level security, so the engine can be pointed at a client's own sales, orders and support data.
Concept build, live demo, 2026
Retail, luxury e-commerce
A luxury watch and jewellery store with no framework and no backend. A concierge reads plain requests, ranks 64 products and explains every pick, across 33 pages from the homepage to a full back office.
33 pages
Concept build, live demo, 2026
SaaS, AI for small businesses
An AI platform concept for small businesses, with a chat assistant demo, an instant quote wizard and six product modules. Underneath sits reusable chat, CRM, payment and portal code that I lift into client projects.
60 routes prerendered as static HTML
Concept build, live demo, 2026
Healthcare, dental clinic
A dental clinic website that answers what nervous patients ask first: who will treat me, what it costs and what results to expect. Before and after sliders, dentist profiles, upfront prices and booking on every screen.
6 services, each with its own page
Tell me what the job is and I will tell you honestly whether it is worth building. Free 30 minute call, no pressure.
Book a 30 minute call
Tell me what the job is and I will tell you honestly whether it is worth building. Free 30 minute call, no pressure.
Open my calendar
