Agentic AI/Live · 10+ DAU
AI-native workspace unifying notes, tasks, projects, and a knowledge graph, with semantic search and a citation-backed multi-agent assistant.
Solo build: product, schema and row-level security, retrieval, the assistant and the GitHub integration.
An AI-native workspace that unifies notes, tasks, projects, and a knowledge graph. Notes, tasks, and projects are nodes with real edges, so the assistant can traverse relationships instead of treating every record as an island.
Local-first UX versus server-side intelligence, multi-agent state sync over Realtime, and predictable LLM costs across multi-step runs.
Retrieval is measured, not assumed. The retrieval eval runs 20 typed queries (paraphrase, exact, mixed) over a 22-document corpus in which 14 documents are deliberate distractors; the first 8-document version scored 100% recall@3 for every strategy, so it measured nothing. Keyword scored 82.5% recall@1 and 92.5% MRR; equal-weight hybrid scored worse (72.5% and 86.7%) until the semantic leg was down-weighted to 0.35. Regression tests assert those floors, so a ranking change that degrades them fails CI alongside lint, typecheck and Vitest.
The eval's semantic leg is a bag-of-words stand-in for a real embedder, so it cannot show what hybrid retrieval is worth; with no API key the product defaults to keyword-only, which is what the numbers support. The assistant answers over retrieved chunks rather than reasoning across the graph, and read-only viewer enforcement for team workspaces is still being swept through.
Local-first UX with server-side intelligence is the right default for developer tools. Unifying storage is easy; unifying context is the actual product.
Tell me what you're building, the constraints you're working with, and where it breaks. I reply within a day.