Financial Date Extraction¶

Client: Viren Rajput, CTO, ScaleXP.

ScaleXP is a financial SaaS platform that automates revenue recognition for hundreds of growing companies. A key step is inferring the service period from free-text invoice lines such as “for services starting next month”, and today this is done by a frontier LLM that is accurate but slow, expensive and not perfectly stable between runs. We would like the team to build a small language model (SLM) - a fine-tuned or distilled specialist - that matches or beats the incumbent on our existing benchmark at around a tenth of the cost, with sub-100ms per-line inference and deterministic output. We can supply roughly 90,000 labelled real-world examples, including around 58,000 high-quality labels derived from customers’ own corrections, together with the evaluation harness we already use to gate production changes. The hard part is the data: relative phrases, ambiguous date formats, backward-looking periods, multiple languages, and lines where no date can honestly be inferred and the model should say so. How to get there - distillation, fine-tuning, base-model choice, quantisation - is open, and the work divides naturally into data preparation, modelling, evaluation and a deployable service. A successful prototype would show whether a specialist model could take on this task in ScaleXP’s live pipeline.