Case study

Precedent

A local-first macOS workspace that turns captured references into usable evidence. AI proposes context; it never silently asserts it.

productmacOSLocal-firstAI productElectronSwift

Live site

Estimated reading time

Key metrics

Context assembly

p95 at depth 1 across 50,000 references

227ms

Capture formats

text, design, image, document, data, CAD

Incorrect-context gate

release fails on a single wrong auto-link

Outcome

Three signals of impact.

Shipped as a signed, notarized macOS app with a self-update feed; releases from 0.1.0 to 0.3.1 over five weeks
Context assembly measured at a 50,000-reference corpus: depth-1 p95 227ms on Apple Silicon, with fail-closed release gates including an incorrect-context rate of zero
58 capture formats, on-device Whisper transcription in a deny-network sandbox and a read-only MCP server so agents get deterministic access with receipts

Role and scope

Solo: product, design system (1,371-line DESIGN.md) and engineering across TypeScript/React, Swift helpers and SQLite service

Constraints

Local-first: no cloud, no telemetry surface; the speech model is not cleared for redistribution, so the beta stays private
One process owns the database; the sandboxed renderer has no Node access and extensions never open the profile store
Themes may change color and atmosphere but never typography, spacing or layout: product ergonomics are non-negotiable

Top decisions

Uncertainty is a product state: ambiguous context stays unassigned instead of being silently filed
Electron accepted conditionally, with a documented exit path and ten non-negotiable security controls
Release quality enforced by benchmarks with hard thresholds, not manual QA

Confidential details redacted where required.

Next step

If you want this outcome, we can scope the engagement quickly.

Contact Work with me