license: mit
language:
- en
pipeline_tag: text-classification
tags:
- dristi
- router
- system-one
- distilbert
- compound-ai
DRISTI V1: The Speed Demon (DistilBERT)
DRISTI V1 is the ultra-fast "System 1" component of the DRISTI Compound AI Routing Engine.
It is designed to act as the high-speed front door to your AI infrastructure. Instead of sending every simple query to an expensive API like GPT-4, this V1 engine handles the routine, easy tasks locally in about 45 milliseconds.
🏗️ How it fits into the Cascade
DRISTI uses a Local Dual-Engine Cascade:
- 👉 V1 (This Model): The frontline router. Handles 80% of routine queries in ~45ms.
- V2 (The Heavy Thinker): The DeBERTa-based safety net. Triggered only if V1 lacks confidence.
When you send a query using the official DRISTI framework, it goes to this V1 model first. If V1 is not confident in its answer (or mathematically detects that the query is out-of-distribution gibberish), it will instantly cascade the query to the V2 heavy model.
📊 Benchmarks
Tested on an NVIDIA RTX 3050 Laptop GPU (4GB VRAM) via the DRISTI router:
| Metric | V1 Fast-Path (This Model) | V2 Heavy-Path | Heavy LLMs (e.g. GPT-4) |
|---|---|---|---|
| Latency | 45.60 ms | 698.57 ms | ~ 1,500.00 ms |
| Throughput | 21.93 QPS | 1.43 QPS | < 1 QPS |
| Cost / 1k | Free (Local) | Free (Local) | ~$20.00 (API) |
🚀 How to Use
To use these weights, you must use the official open-source DRISTI inference engine, which handles the complex routing logic, dual-engine memory management, and FastAPI deployment.
- Clone the main repository:
git clone https://github.com/SWARAJCHATTARAJ/DRISTI.git cd DRISTI - Download the model weights (
pytorch_model.bin) from this page and drop them into yourcheckpoints/directory asdristi_v041_best.pt. - Start the dual-engine router:
python -m inference.dual_engine
(Note: We've included dristi_model.py here for reference if you're using advanced HF loaders, but we highly recommend using the official DRISTI repo for actual deployments to get the full dual-engine benefit.)