MY MOM ANUPAM SINGH MADE A DEVPOST ACCOUNBT AND LET ME, RONIT SINGH, PARTICPATE
Inspiration
As I've gotten more caught up in the developer and AI world, I began to become very interested in the progress of Frontier-Level AI models from OpenAI, Anthropic, and more recently labs like DeepSeek and Moonshot for programming and general use tasks. A big part of this recent interest is monitoring model releases and their performance on AI benchmarks like SWEBenchPro, DeepSWE, and Artificial Analysis. But, I felt a big disconnect from these arbitrary scores I saw on these benchmarks and the performance I saw in my own work. So, I made Ariadne.
What it does
Ariadne is a AI testing and Benchmarking tool that is runs on your code and grades on what YOU value.
How we built it
Ariadne is a local-first Node.js and TypeScript CLI. I used Codex with GPT-5.6 throughout Build Week for the following: exploring the existing architecture and iterating on it, testing edge cases, reviewing changes, making and planning new features, and testing Ariadne against disposable repositories.
Commander wires thin commands, Zod and YAML validate inputs, Execa manages processes, Git provides attributable change evidence, and Ink/React powers the TUI. There is no hosted backend or hidden database.
Challenges we ran into
The hardest part was executing the vision for Ariadne, I had a general Idea of what i wanted to do but I had to stumble along the way to really find the niche this product needed to target and what it was meant to function as for devs.
Another thing was the TUI, codex is still not the best at designing TUIs, so balancing info and good design for the TUIs using my own taste that codex doesn't have took me a while.
Accomplishments that we're proud of
I'm happy I grew this project from a simple runnable tool to a fully built CLI tool that has guardrails for safety and varying degrees of verbosity so anyone can use it how they want I am are especially proud of the safety details: ignored-file checks can still catch paths such as .env, read-only tasks are verified against repository mutations, promotion revalidates repository identity and cleanliness under a lock, and conflicts are preflighted and rolled back rather than left in the user's checkout. \
What we learned
I learned what devs actually need in their tools, as Codex flagged a lot of things along the way that I wouldn't have thought about.
I also learned how to more effectively use AI in my day-to-day dev work
Built With
- ai
- benchmark
- cli
- codex
- typescript
Log in or sign up for Devpost to join the conversation.