Back to projects
Open Source / AI Engineering / 2026
ArchDrill
Open-source “LeetCode for system design”: problems with hard numeric constraints, a React Flow design canvas, a deterministic load simulator, and BYOK AI grading against hidden rubrics — all running in the browser with no backend.
Case studies
Case 01
AI Grading Without a Backend
Designed a browser-only grading pipeline: user diagrams are serialized to text and scored by Anthropic or OpenAI models against hidden rubrics, with keys living exclusively in localStorage.
- Problem
- Interview-quality feedback usually requires a server, accounts, and stored credentials — friction that kills a practice tool.
- System focus
- Diagram serialization, rubric design, deterministic client-side load simulation, and chaos scenario stress tests.
- Outcome
- A serious practice loop — design, simulate, break, get graded — with zero infrastructure and zero telemetry.
Highlights
- React Flow canvas with typed components, replica counts, annotations, and undo/redo
- Deterministic load simulator reporting served RPS, latency, availability, and per-component utilization
- Chaos scenarios (AZ outage, cache cold start, 10x spike) stress-test designs before grading
- BYOK AI grading with per-dimension scores, issues paired with fixes, and attempt history
- One-click AI generation of new problems with rubric, reference solution, and diagram
React FlowZustandAnthropic APIOpenAI API