InscribeAI · posted · checked
Data Scientist
AWSDjangoGoPythonPytorchOn-site · Dublin
Who can apply
On-site · Dublin
Location: Dublin
ABOUT INSCRIBE
At Inscribe, we're on a mission to catch fraud with confidence. In 2017, we set out to build the first document fraud detection product. Today, we process millions of applications each year and work with many of the leading fintechs and banks across the U.S. We believe the game of cat-and-mouse in fraud requires a fundamentally new approach. With care for our customers, attention to detail, and grounded execution, we’re reimagining how fraud is detected and prevented.
Inscribe is trusted by leading banks and fintechs, including Coinbase, Plaid, Bluevine, and Ramp. Inscribe is backed by Threshold Ventures, Uncork Capital, Y Combinator, Crosslink Capital, and Foundry, and is headquartered in San Francisco, CA.
THE OPPORTUNITY
Fighting fraud is a continuous battle. We’re hiring a Research Scientist to develop our fraud detection capabilities by investigating new methods, improving the ones we have and, most importantly, turning research ideas into production algorithms.
The documents that pass through Inscribe are many and varied in content and format. Images are a particularly tricky file type and we are always looking for ways to extend our catch rate on them. We look for someone eager to quickly test prototypes, solve issues as they come, learn on the go.
Research at Inscribe is not a siloed function but it sits at the heart of our tech team and collaborates very strictly with engineering: anyone is empowered to come up with ideas and try them, fraud fighting is a shared goal. Research informs what we do, rigorously tests ideas, quickly discards those that don’t work and communicates findings in an approachable way.
RESPONSIBILITIES
Own research problems in document fraud detection from start to finish
Help build detectors for fraud detection algorithms, with a special focus on image file types, such as edited or tampered images, JPEG manipulation and AI-generated documents. Combine classical forensics, deep learning and VLMs.
Design rigorous evaluation: sampling strategies to estimate recall on unlabelled production traffic, ground-truth labelling, precision/recall trade-offs at low fraud rates, backtesting and shadow experiments.
Find and sort failure modes by digging into real customer documents, and turn them into labelled datasets and new detection work.
Work on LLM-based fraud detection: model and tool evaluations, agent and prompt experiments, customer-level analysis and reactive fixes.