This post was written as part of the Kiroween 2026 hackathon.

Most health-tech tools assume the person taking the medication has a smartphone, an account and the patience to tap through an app every morning and evening. In real life, that breaks down fast in elder care. Parents and grandparents do far better with a printed sheet on the fridge and a pen.

On the other hand, snapping a photo of a paper chart and sending it to a Vision LLM creates a worse problem: it is slow, sends private medical data to external servers and risks hallucinating faint or messy pen marks. In medication tracking, guessing a missed dose as "taken" is a safety hazard.

For the Kiroween 2026 hackathon, I built PaperBridge: an open-source, offline-first optical skeleton that turns plain printed paper into a fast, deterministic digital interface, powered by Kiro's full developer workflow.

Elder Care First: Paper as the UI, AI Only Where It Belongs

PaperBridge separates optical reading from AI reasoning:

  1. Deterministic Optical Core (src/core/): Runs 100% client-side in the browser in 28.5 ms median with zero backend and zero Vision LLM calls. It locates 4 corner fiducials, corrects perspective and classifies every bubble locally.
  2. Honest Uncertainty + Human-in-the-Loop: Ambiguous or faint pen marks are never guessed. They are flagged as unclear and the caregiver can click the bubble to verify it as Filled or Blank. The raw optical scan stays immutable, while a "Manually verified (N)" audit badge is stamped on the result.
  3. Real Caregiver Workflows (src/apps/elder-care/): Caregivers customize the patient name, week start date and medication or habit labels before printing a PDF card (bubble coordinates stay identical so the decoder needs zero changes). Scanning tracks adherence, flags missed doses, saves multi-week trends locally and copies a clean Markdown Doctor Visit Summary in one click.
  4. True Offline Resilience: Packaged as an installable PWA with a service worker precaching the app shell, PDF generator and sample cards, so printing, webcam scanning and saving history work even with Wi-Fi turned off.

Built as a Reusable Skeleton

While Elder Care Bridge is the flagship workflow, PaperBridge is architected as a reusable optical skeleton (src/core/ + src/shell/). Any developer can run npm run create-app <slug> "<Display Name>" to scaffold a brand-new paper-based application (such as clinic intake forms or field checklists) with a valid card layout, pure interpreter, <ScanWorkbench> UI and passing property tests in under a minute.

How Kiro Powered the Build

Instead of using Kiro just for one-off code generation, I used all five pillars of the Kiro workflow:

  • Spec-Driven Development (.kiro/specs/paperbridge/): Every core invariant, accessibility requirement (WCAG 2.2 AA/AAA) and workflow phase was tracked across requirements, design and tasks. That discipline produced 149 passing tests, including fast-check property-based tests that verify decoding guarantees.
  • Steering Docs (.kiro/steering/): Custom steering rules lock down the separation between src/core/ and src/apps/ so new features never weaken optical thresholds or mutate raw scan results.
  • Agent Hooks (.kiro/hooks/): Three automated hooks keep the core invariants, sample generation and UI checks aligned as the codebase evolves.
  • Custom Layout Validator MCP Server (mcp/ + .kiro/settings/mcp.json): When designing a new paper card, LLMs can easily place bubbles too close together or overlap corner fiducials. I built a local MCP server exposing validate_card_layout and list_registered_apps so Kiro validates card geometry deterministically before printing.
  • Kiro Crew App (PaperBridge Care): In Elder Care Bridge, clicking Copy for Kiro Crew copies the verified scan JSON (including the "Manually verified" audit trail). Pasting it into PaperBridge Care inside Kiro Crew launches the care-auditor agent with the adherence-analysis skill to draft a warm, accurate caregiver check-in.

What I Learned

The biggest takeaway from building PaperBridge for Kiroween 2026 was knowing where not to use an LLM. Keeping the optical layer 100% deterministic, local and property-tested made the elder care workflow trustworthy, while using Kiro's specs, MCP validation and Kiro Crew agents turned a single-scan prototype into a complete, extensible platform.

Try the live offline demo at paperbridge-acvw.onrender.com or fork the repo to build your own paper-based app!