Within hours of the earthquake that struck Turkey, a group of programmers opened a Discord server and began building afetharita — literally “disaster map” — an application designed to help search and rescue teams locate survivors. The problem was immediate and overwhelming: survivors were posting screenshots of texts containing their addresses and urgent needs directly to social media. That unstructured flood of tweets and images needed to become structured, actionable data, fast.
The team’s technical approach was pragmatic under pressure. For the screenshot pipeline, they landed on easyocr for text extraction, then fed the output through a transformers-based fine-tuned NER model built on bert-base-turkish-cased. That model extracted names, phone numbers, and addresses, which were then sent to a geocoding API to produce coordinates for the front-end map. Hugging Face infrastructure proved essential: they hosted the app on Spaces with a GPU grant, used the Inference API to serve models without managing Docker images or CI/CD pipelines, and got extra replicas to handle traffic spikes without downtime.
A second challenge emerged when volunteers asked for a model that could classify what survivors needed — shelter, food, logistics — from raw tweets. The team first experimented with zero-shot NLI models like xlm-roberta-large-xnli and a Turkish ConvBERT variant, which let them infer directly with candidate labels without labeled data. Few-shot prompting of closed-source models was also on the table. But generative models risked hallucinating labels, so they ultimately fine-tuned their own classifier. The turnaround was astonishingly fast: roughly three minutes to fine-tune BERT’s classification head on a single GPU.
This effort echoes other crisis-response ML deployments — like the use of satellite imagery and NLP after Hurricane Maria — where speed matters more than elegance. The difference here is how much modern tooling lowered the barrier: a distributed team of volunteers, coordinated through pull requests and a shared model registry, shipped a working rescue tool in the immediate aftermath of a disaster. The leaderboard they built from model card metadata meant production decisions could happen in real time, not in a planning meeting. It’s a model for what volunteer engineering can do when the stack gets out of the way.
