NoHaram DNS
AI/MLAI-driven domain classification and filtering with sub-15ms latency across 2.1+ million domains using a fine-tuned DistilBERT model.
- Year
- 2025
- Role
- ML and systems
- Stack
- Python, DistilBERT, Redis, CoreDNS
What it does
Fine-tuned DistilBERT to classify domains in real time, then wired the model into a CoreDNS filtering layer. Redis sits in front of the model so repeat lookups never pay the full inference cost.
It runs under 15ms per classification across a corpus of more than 2.1 million domains.
Why latency is the product
This is an inference path sitting in a hot loop. DNS resolution blocks every page load, so classification time is not a nice-to-have. The lambda-style architecture (cache first, model on miss, write-back) is what makes the accuracy usable at the edge of a request.