Inspiration
Every QA team maintains hundreds of test cases with steps. Click this. Fill that. Press submit. But when the UI changes — and it always changes — someone has to update all of them. Nobody does. Tests become outdated, teams lose confidence, and regressions slip through.
The problem isn't discipline. The problem is the abstraction.
Step-based test cases turn testers into script-readers. They describe how to click, not what to verify. They duplicate the same "login" flow across fifty test cases. They're written by humans but not natural to humans — because users don't think in steps. They think in goals.
What it does
Telos replaces test cases with a knowledge graph where every node is a user goal — "buy a thing", "leave a review", "return an item" — connected by dependencies.
Each node contains a semantic description: what information is needed, what validation applies, what success looks like. Not UI steps. Business rules.
From this graph, Telos:
- Generates test coverage automatically via graph traversal — every possible user flow, nothing missed
- Synthesizes tests on demand via AI — exploratory brief for senior testers, step-by-step for juniors, same node, different format
- Exports a full test suite as an Excel artifact for release notes and compliance
- Enables targeted regression — change one node, find all affected flows instantly
The graph is the truth. Tests are ephemeral artifacts generated when needed.
How we built it
Backend: Python + FastAPI + networkx for graph storage and traversal, Anthropic Claude API for AI synthesis
Frontend: React + React Flow for the graph editor, dagre for automatic layout, TipTap for rich semantic node descriptions
AI layer: Each generate request passes the node's goal, semantic description, predecessor nodes, and successor nodes as context. The model synthesizes either an exploratory brief or a step-by-step test case depending on the selected mode.
Challenges we ran into
The hardest design decision was the right level of abstraction for a node. Too granular ("fill email field") and you rebuild the step-based problem. Too broad ("use the app") and the graph loses meaning. The answer was JTBD-thinking: a node is one complete thing a user wants to accomplish, described in terms of information and validation rules — not UI interactions.
Accomplishments we're proud of
The graph traversal generating real coverage paths felt like the moment the concept clicked. It's one thing to say
