Grant consultants spend days manually searching for government grants across dozens of sources: federal databases, state portals, listing pages, the open web. Then they repeat the whole process for the next client, cross-referencing eligibility criteria against business profiles by hand. The high-value grants get buried under hundreds of irrelevant ones.

I saw a manual, expensive process that wanted to be a product. So I built it, end to end.

Days of searching, down to one scan for about a dollar. An operator creates a client profile once, triggers a scan across five discovery layers, and gets a ranked list of matched grants with plain-language rationale and direct links. I validated the opportunity with working consultants and tested with real start-ups before shipping working software. Not a prototype, not a deck. Software people use.

I owned everything: problem definition, UX architecture, product spec, scoring framework, and the build itself. Every decision was shaped by a constraint most designers never face: each API call costs real money. That constraint produced the three design decisions the product stands on.

Discovery is decoupled from scoring. Operators control spend. They can cast a wide net cheaply, then choose what's worth the expensive analysis. Cost transparency is a feature, not a footnote.

Users see the AI thinking. Pause gates between pipeline layers turn a black box into a process you can watch and trust. When an agent is working through hundreds of grants, showing its reasoning is the difference between a tool consultants rely on and one they double-check.

Triage runs on actions, not scores. Results lead with Apply Now, Watch List and Skip, because consultants prioritise by next step, not by a number. Underneath sits two-pass AI scoring, Haiku for cheap bulk filtering and Sonnet for deep refinement of top candidates. It's a product design decision disguised as architecture: nobody ever waits on expensive processing for grants that clearly don't match.

If you're trying to turn a manual process into an AI product, this is the work I do end-to-end. Let's talk.