Confidential · Capability Reference
Documents into data
you can trust.
Codeaza Technologies builds production document-intelligence systems: high-fidelity extraction from complex published documents, an intelligence layer that scores and normalises every record, and independent cross-validation that catches defects before they ever reach the consumer.
When the data feeds machines, "mostly right" is a failure.
A capital-markets client draws reference data from a large, ever-changing library of complex published schedules. The output feeds automated decision systems and is independently diffed against source, so the bar is trustworthy-without-checking, not best-effort.
row
A naïve "LLM reads the PDF" produces plausible output with no measure of confidence and no way to catch the one wrong row in ten thousand before it ships. That gap is the whole engineering problem.
Three layers: extract, make intelligent, cross-validate.
Each layer has one job and a clear guarantee. Extraction stays faithful to source; the intelligence layer makes meaning explicit and measured; validation re-derives independently and gates the hand-off.
What each layer actually does.
LLM-assisted extraction over heterogeneous PDF/HTML, chunked so every value traces to a specific source region. Resolves multi-column "Fee/(Rebate)" split headers, footnote-driven overrides, mixed currencies (USD/EUR/GBP/ZAR/NOK), bps and compound expressions — and tracks each revision for change detection.
Raw text is decomposed into a typed, enumerated schema — a signed numeric value with explicit direction (charge / rebate / credit), currency and unit, not free text. A separate scoring pass grades every record for correctness against source and every section for completeness. The scores travel with the data.
An independent pass re-derives the answer from source and flags any disagreement — a value whose sign contradicts its label, a rebate stored as a positive charge — as a defect candidate. Constraint validators enforce typed domains and internal consistency; a pre-delivery audit gates the hand-off.
The value isn't "we got the data out." It's the layer on top: a confidence signal on every record and an independent check against source that turns a best-effort extraction into a delivery you can gate on.
The engineering that makes it trustworthy.
- Extraction is probabilistic — the same schedule can be read a dozen ways, and one dropped parenthesis flips a rebate into a charge.
- Meaning is buried in layout: split headers, footnotes, tape/price-band dimensions, mixed currencies and units.
- The consumer diffs our output against the original — so values must stay verbatim-faithful while their economic meaning is made explicit.
- Probabilistic in, deterministic out — LLM extraction wrapped in deterministic validators + regression evals, so quality is measured, not hoped for.
- Dual-axis confidence — correctness (per record vs source) and completeness (did we miss a stated line?) scored separately; two failure modes, two signals.
- Independent cross-check — validation re-derives from source on a separate path, catching the extractor's blind spots instead of rubber-stamping them.
- Sibling sweeps — each new defect class triggers a full back-catalogue sweep, so the second instance is found by us, not the consumer.
Every record is scored, constrained, and gated before it ships.
Correctness — does the record faithfully match its source region (amount, sign, participant, tier, classification)? Completeness — of everything the source states in a section, how much did we capture? Both stored on the record.
Validator families run at write and pre-delivery: sign ↔ direction consistency · currency ∈ enumerated domain · parsed value ↔ raw text reconciliation · provenance (every row maps to a source chunk) · completeness reconcile (extracted ≥ stated).
Green — high confidence, zero validator failures → auto-deliver. Amber — deliver with a flagged-rows manifest. Red — held for human review; nothing ambiguous ships blind.
The consumer isn't asked to trust a black box. They're handed a measured, verifiable data product — with the confidence to gate their own systems on it.
Measured quality, at production scale.
history
Values traceable line-by-line to source and survive the consumer's own diff. Defects surfaced by our audit — and swept for siblings — before the client sees them.
The domain changes. The discipline is the asset.
Faithful extraction, an intelligence layer that scores and normalises, and independent cross-validation that catches defects before the consumer does — the same method applies wherever documents must become trusted structured data: contracts, regulatory filings, specifications, catalogues, rate cards. Measured, verifiable, prevention-first.