🤓 Latest Submissions
Velora
Velora is a local-first AI inference agent built for the AMD Hackathon Track 1 scoring track. It reads a task list, classifies each prompt into one of eight categories—factual, math, sentiment, summarisation, NER, debugging, logic, or code generation—and answers every task with a small local model running inside a Docker container. A confidence-cascaded router decides whether the local answer is solid or should be escalated to Fireworks API models in hybrid mode. In zero-token mode, Velora stays entirely local and never calls a remote API. Each category has a dedicated solver with built-in verification: independent Python programs vote on math answers, generated code is compiled and fuzzed against test cases, logic puzzles are solved by brute-force enumeration, and text tasks use structured prompting and strict format checks. A watchdog guarantees a complete results.json before the hard deadline, while the flat, dependency-light codebase is validated against the local benchmark harness before each release.
13 Jul 2026
