Small thing I'm proud of in Nexus: the glow behind the search bar follows your cursor, but the code isn't running an animation loop nonstop. It only asks for a new frame while it's still catching up, then stops. Nothing spins when nothing's moving.
When you open something in Nexus, the icon and title don't just swap to a new screen, they glide into place. Same element, new spot. Small trick, but it makes the jump feel connected instead of two screens stitched together.
One rule I kept coming back to while building Nexus: no navbar, no hero section, no buttons floating around. If you want to do something, you just type it. Sounds small, but it changes what counts as a feature worth building.
A command palette doesn't need an LLM, it needs to be instant. Built a dependency free fuzzy matcher: bonuses for consecutive runs and word boundary starts, similar to fzf. Ruled out semantic search here, deterministic beats a network call on every keystroke.
Worth naming a real limitation: Nexus only matches literal characters, not meaning. "Doc Arjun sent" won't match unless those words appear. Semantic search needs embeddings and a vector index, a different latency budget. Not adding it without justifying the cost.
Spent time on how selection moves between grouped results, not just whether it does. A background swap reads as "the state changed," not "I am here." The selection indicator is one element that animates to the active row, so your eyes track motion instead of noticing a cut.
@pu4pr exactly. i'd rather keep the hot path deterministic and bring in semantic search only when fuzzy matching falls short. easier to debug, cheaper to run, and way more predictable
Starting a build in public series. First up is Nexus, a command palette concept. The real problem isn't the UI, it's ranking: deciding what a 3 character query means without calling a model on every keystroke. More on the implementation this week.
Starting a build in public series. First up is Nexus, a command palette concept. The real problem isn't the UI, it's ranking: deciding what a 3 character query means without calling a model on every keystroke. More on the implementation this week.
GM GM
Another Win!
Didn't knew my college could pull BlackRock,JP MorganChase and other goated companies as panelists and judge projects
Presented one of my projects and Won😁👍
Didnt knew we had to present in front of such respected personalities and this big of a crowd till last night
@Utsav3Garg our speedrun will be remembered👽
@kushagr09 @CyreneAIFlows @CyreneAI Medical sector is really the one that needs to be revolutionised, I can't really afford to pay so much for the consultation cost