Document Data Extractor Service
Extraction and classification services built from scratch — deep learning over unstructured text, delivered to production by a team of six I led and mentored.
- Role
- Design owner · led team of 6
- Scope
- Built from scratch
The problem
Unstructured documents carrying the data a business actually runs on, and no way to get at it except by reading them. Extracting it reliably is a machine learning problem; exposing it as a service other systems can depend on is an engineering one. Most efforts solve the first, demo well, and stop — which leaves the organisation with a proof that it is possible and nothing it can use.
The approach
Built the extraction and classification services from scratch and led the team that delivered them:
- Deep learning over unstructured text, with embedding-based representations doing the work that keyword and rule-based matching could not.
- Document processing microservices, designed so extraction and classification were separately deployable and could scale independently of each other.
- ETL into Synapse and the data lake, so extracted output landed somewhere downstream systems could actually consume rather than in a service’s own storage.
- A team of six led and mentored through delivery, with me owning the design and carrying it to production instead of handing over a prototype.
The outcome
Production services, built from the ground up and delivered — which is the distinction that matters here, since the hard part of extraction is never making it work once. It is making it work every time, on documents nobody anticipated.