.png&w=256&q=75)
🤝 Top Collaborators
🤓 Latest Submissions
Adaptive Cost-aware LLM router
RouteZero is a hybrid token-efficient AI routing agent built for the AMD Developer Hackathon ACT-II Track 1. It solves the core enterprise AI problem of minimizing remote API token costs without sacrificing answer quality. Every incoming prompt is processed through a multi-stage pipeline: first hitting a ChromaDB semantic memory cache (cosine similarity ≥0.92) for zero-token responses on repeated queries, then passing through a custom Semantic Heuristic Router that classifies prompts across 8 capability domains using sentence-transformer embeddings and weighted heuristic signals. Simple tasks (factual knowledge, math, sentiment, summarization, NER) are handled locally by a Qwen2.5-1.5B 4-bit quantized GGUF model running on CPU at zero Fireworks token cost. Complex tasks (code generation, debugging, logical reasoning) are escalated to a frontier model via Fireworks AI. A deterministic verification layer catches low-quality local responses and triggers automatic fallback to remote inference. The entire pipeline is orchestrated via LangGraph state machine, containerized for linux/amd64, and optimized for the 4GB RAM / 2 vCPU grading environment. RouteZero demonstrates that intelligent routing — not raw compute — is the key to scalable, cost-efficient AI in production.
13 Jul 2026
