Blockradar, an AI sprint facilitator, built end-to-end.
A self-built product prototype: a bilingual tool that captures, triages and prioritizes sprint blockers in ERP implementation projects, with an optional two-stage AI backend on Cloudflare Workers. Designed, built and deployed solo, from concept to running system.
The idea
In ERP implementation projects (IFS Cloud, Dynamics 365), sprints rarely fail for lack of work, they fail because of blockers: a pending decision, a waiting department, a dependency on another team. These impediments surface verbally in standups, disappear, and resurface. Blockradar makes them systematic: every blocker is captured, triaged, prioritized and tracked on a single radar board.
The project combines three things I work with daily: agile delivery (as a certified Scrum Master), ERP project reality (20+ years IFS), and modern full-stack development with AI integration.
▸ Open the live demo — BlockradarWhat it does
Automatic triage
Each blocker is classified on entry: category (decision / dependency / technical / external), impact on the sprint goal (high / medium / low, with automatic escalation of aging blockers), a concrete next step, and the right role to escalate to. Every suggestion is editable, the system proposes, the facilitator keeps the last word.
Priority radar board
Blockers are sorted by criticality (impact × age), with live statistics, filters, a status workflow (open / resolved) and visible ownership, who raised what, and how long it has been open.
Bilingual by design
The full interface, including the triage output switches between German and English at a click, built for DACH project teams.
Runs anywhere, no server required
The frontend is a single self-contained HTML file: browser persistence, JSON export/import for portability, and a built-in rule-based triage engine, so the demo runs standalone, with zero infrastructure.
Optional two-stage AI backend
When connected to its Cloudflare Worker backend, triage is performed by real AI using a tiered model strategy: a fast, inexpensive model (Haiku) extracts and classifies; a stronger model (Sonnet) then reasons about impact, formulates the next step and the escalation, in the user's language. The frontend enhances progressively: instant local triage first, AI refinement seconds later. If the backend is unreachable, it falls back cleanly to the built-in logic.
How a blocker is triaged
The architecture separates the instant local path from the optional AI path, the tool is always responsive, and AI adds depth when available. The API key lives server-side as a Cloudflare secret and never reaches the browser.
Engineering decisions worth noting
- Tiered model strategy a cheap, fast model for extraction and a stronger model only for the reasoning step keeps AI cost per blocker minimal without sacrificing output quality.
- Key security by architecture the AI key exists only as a Cloudflare secret in the Worker; the browser never sees it.
- Progressive enhancement the UI never waits for the AI: local triage renders instantly, AI refinement lands seconds later with a visible status.
- Human in control every AI suggestion (category, impact, escalation) is editable; manual entries are never overwritten by automation.
- Zero-infrastructure demo mode a single HTML file with browser persistence and JSON export/import runs the full experience without any backend.
Tech stack
A complete product cycle executed solo: concept, UI/UX design, a working frontend, a deployed serverless AI backend with a cost-aware two-stage model strategy, and honest product judgment — built as a prototype to demonstrate exactly the combination I bring to client work: ERP domain knowledge, agile practice, and modern full-stack engineering with AI.
Need a tool like this built?
Custom applications with AI integration — from concept to deployed system.
Start a project →