Risk Radar — AI Project Risk & Decision Engine
Home → Projects → Risk Radar
AI PROJECT MANAGEMENT · RISK ANALYTICS · NYU SPS × ASANA
I built Risk Radar for the NYU SPS × Asana Challenge 2026 as a project intelligence tool that helps managers identify risks and make important decisions before work begins or issues become visible in status updates. The Streamlit application turns a project brief and team roster into a skill-matched plan, a project health score, and a prioritised decision agenda.
Risk Radar is designed to support project planning and risk awareness. Its recommendations are intended to inform managerial decisions, not replace human judgement.
System
AI project risk and decision engine
Stack
Python · Streamlit · Anthropic Claude Haiku 4.5
Focus
Project planning, workload balancing, risk scoring, and decision support.
The Problem
Project management tools often make it easier to track tasks and report issues after they appear. However, managers also need to identify risks before execution, allocate work according to team skills and capacity, and decide which unresolved questions require immediate attention. Risk Radar addresses this gap by bringing planning, risk scoring, workload visibility, and decision support into one workflow.
What I Did
I built Risk Radar as a single-file Streamlit application powered by Anthropic’s Claude. I designed four connected modules: a Project Planner that generates a structured task plan and schedule, a Dashboard that calculates project health using local risk rules, a Decision Engine that ranks important managerial decisions, and a Chat interface that answers questions using the current project state.
Project Planner
I implemented a planning workflow that accepts project scope, goals, team details, and deadlines, with an option to extract structured fields from an uploaded brief. It generates a 12–15-task plan with phases, skill-matched assignees, a schedule, and the ten most critical pre-project decisions. A keyword-based skill map and workload-balancing logic help distribute tasks across the team, with a target of keeping any one member’s share at approximately 40% or less.
Risk Dashboard
I implemented a local project health and risk-scoring engine that produces a score from 0 to 100. It identifies risk drivers such as overdue, unassigned, and stalled tasks, as well as high-complexity tasks without deadlines. The dashboard also surfaces individual workloads and potential overload warnings. These calculations run in Python without requiring an API call.
Decision Engine
I built a decision-support workflow that analyses an uploaded project brief and surfaces the ten highest-impact decisions for the manager. Each decision includes available options, a recommended course of action, and supporting rationale. The application caches results after the first run to reduce repeated token usage.
Context-Aware Chat
I implemented a chat interface for questions about project risks, team capacity, estimated hours, timelines, and individual tasks. By injecting the current project state into the conversation, the application grounds its responses in the available plan rather than relying solely on generic project-management advice.
Engineering Decisions
I kept risk classification and health calculations in pure Python so the dashboard can work without API calls. I designed keyword-based skill categories covering backend, frontend, design, QA, machine learning, DevOps, and project management, with an LLM-assisted assignment process supported by a local fallback. I used brief compression, content-hash-based response caching, history trimming, and a live token meter to control model usage. I implemented a four-stage JSON repair process for common formatting issues and split planning into two calls—tasks and schedule, then decisions—to reduce truncation risk. Uploaded documents are parsed into a structured schema; this is not a retrieval-augmented generation system and does not use a vector store, chunking, or a retrieval step.
Impact
Risk Radar brings task planning, workload balancing, local risk scoring, and AI-assisted decision support into a single interface. Its combination of deterministic Python rules and LLM-powered workflows is designed to help managers identify potential issues earlier, understand the factors contributing to project risk, and prioritise decisions before execution.
What it Demonstrates
Risk Radar demonstrates how AI can complement conventional project-management workflows. I combined local risk logic, skill-aware task assignment, structured LLM outputs, token-conscious orchestration, and context-aware interaction to build a practical decision-support application.
Python · Streamlit · Anthropic Claude Haiku 4.5 · pypdf · Asana REST API