Technology, data and technical reports

Home Cases Decades of technical reports turned into data an…

Ship Shift · Cases

Decades of technical reports turned into data and predictive models

Last updated: 2026-11-17

We turned the historical report archive — including scans and handwriting — into a structured database and scoped predictive models on top.

Geoscience company in energy (UK). Confidential — name not published.

The challenge

Technical well data lived across PDFs spanning decades: modern typography, low-quality scans and a non-trivial share of handwritten notes. The geoscience team hand-copied critical fields into internal databases to analyse trends. It was slow, expensive and impossible to scale at the pace the business needed.

Without a clean database with document→field lineage, any predictive model was fragile: you trained on incomplete tables or inconsistent campaigns. The risk was not only typing cost — it was deciding on partial evidence.

The client needed two things in order: first, reliable extraction into an agreed schema; then scoped models only where historical coverage allowed. Reversing that order («model now») would have amplified error.

What we did / The solution

We built a LangGraph agent pipeline aimed at closed document types, with human review on critical fields and PostgreSQL load with traceability. Predictive models arrived only after validating coverage and history.

  • Classify useful pages vs. noise (covers, irrelevant annexes, duplicates).
  • Extract fields into the schema agreed with the client's technical team.
  • Vision + LLM for scanned PDFs; field-type rules and validators.
  • Human review queue for high-impact or low-confidence fields.
  • Database load with document→page→field lineage and schema versions.
  • Scoped predictive models on subsets with enough history — not magic AI over gaps.

We worked under sector confidentiality: we do not publish the name or identifiable screenshots. The technical pattern (archive→base→model) is reusable and is what we document here.

Every critical field had an owner on the client side. We agreed confidence thresholds by field type and what to do when a document contradicts a prior version.

Results

Authorised qualitative outcome: the team stopped retyping the archive to feed analysis; the structured database became the starting point for models and reporting. Analysts spend time interpreting, not transcribing.

We do not invent global savings percentages or accuracy claims without a public client measurement protocol. Internally, field-level precision was measured per document type before widening scope — that discipline was part of the outcome.

The compound effect shows when the next report no longer starts from zero: schema, validators and lineage are ready, and the marginal cost of a new batch stays down.

How we rolled it out

01

Closed document pilot

One report type and a narrow schema. Measured field-level extraction (including scans) and set human-review thresholds.

02

Pipeline industrialisation

LangGraph agents, queues, lineage and DB load. Trained the exception validators. Expanded to neighbouring types only with stable metrics.

03

Scoped models

With clean history, trained/validated predictive models on business-prioritised use cases. Governance: the model does not run where coverage is missing.

Lessons & who it fits

Field-level accuracy rules. A narrow, well-measured schema beats «extract everything» without a data owner. Handwriting has a ceiling — quality was measured before promising full coverage.

It fits energy, infrastructure and engineering with dense historical technical archives. It fits less if documents already birth structured in one system, or if nobody can validate critical fields in the pilot.

FAQ

Why confidential?
Sector and contract; the archive→base→model technical pattern is publishable.
Handwriting?
With limits. Quality measured per type before widening — no miracle OCR promises.
On-prem?
Possible per client policy and data restrictions.
Reusable elsewhere?
Yes — the pattern repeats in any dense technical archive.
Pilot length?
Weeks on a closed document type with an agreed schema.
Do models replace the expert?
No. They scope predictions where history exists; the expert validates and decides.

Is your history still trapped in PDFs?

Describe document types and volume. We'll propose an extraction pilot with metrics.

Let's talk