Built specifically for Track 1 of the AMD Developer Hackathon: ACT II, Token-Slasher is a lightweight Python-based orchestration agent optimized for high-efficiency batch evaluations. Instead of relying on heavy algorithmic routing layers that consume extra compute, this agent employs a "smart-minimalist" fallback strategy. It dynamically parses incoming allowed model lists to target the lowest-cost open architectures available on the Fireworks AI platform (such as Llama-3-8B). To drive token consumption down to the absolute floor, it intercepts user queries and pairs them with a hyper-focused system directive that eliminates introductory phrases, explanations, and unnecessary text context. Combined with a hard execution ceiling on output tokens and forced deterministic sampling, the agent ensures highly accurate task completion with a near-zero token overhead footprint.
Category tags: