Aelix Echo · Document Intelligence
Documents that
read themselves.
For ops teams at insurers, banks, and manufacturers drowning in PDFs. We extract the fields that matter from every document, with a confidence score on each field and an audit hash on every page.
Agentic AI · Document agents · Now live
Agents that read, validate, and reconcile. A human signs the write-back.
Document intelligence is not just OCR. It is deciding what each field means, validating it against a source of truth, and reconciling it back into the system that owns it. Aelix Agents do that work end to end. They capture, extract, validate, and propose a reconciliation. A credentialed human approves anything that mutates a system of record.
Document Capture Agent.
Ingests PDFs, scans, and phone photos. Classifies the document type, runs OCR through a pluggable provider interface, and packages the extraction with per-field confidence.
Field Extraction Agent.
Pulls structured line items from messy documents (invoices, BOLs, claims, statements) and validates them against schema. Confidence below threshold escalates to a reviewer with the agent's draft pre-filled.
Validation & Match Agent.
Cross-checks extracted values against the matching source record (PO, ASN, claim file, ledger entry). Flags mismatches as drift records with the specific field and the source values cited.
Reconciliation Agent.
Drafts the write-back. Auto-applies only within tenant-set thresholds (low value, clean match, reversible). Anything ambiguous waits for a human.
Trust architecture
Citation or strip.
Every extracted value cites a specific page and bounding box. Every reconciliation cites the source record it matched against. Uncited output is discarded before it reaches a human.
Fail closed, never fabricate.
If OCR is unavailable, a labeled deterministic fallback keeps the pipeline green rather than failing the document. If the validator cannot resolve, the agent escalates. Never invents a match.
Audit-logged, every action.
Every capture, extraction, validation, and write-back lands in the same tamper-evident audit log a human action does. The agent cannot forge or skip the audit.
Human-approved on the write.
Read tools run automatically. Write tools (post a payable, update a ledger, close a record) are dedicated, gated, and require either tenant policy approval or a human token.
In production · Customs operations
Customs paperwork, read by the agent chain.
A customs import and export team.
Two platforms: one for import documentation and one for export documentation, including the paperwork needed to clear shipments. A chain of AI agents reads and checks each document, people review only the flagged items, and approved data moves on through a process automation layer.
Read the case studyScope
- Documents
- Import and export documentation, including Bill of Summary and Export Summary
- Platforms
- One for import, one for export
- Automation layer
- Microsoft Power Apps and Power Automate
The thesis
A document is just structured data waiting for someone to type it in.
The cost of that typing, in time, in errors, in re-keys, in delayed downstream work, is the cost most operations teams have stopped measuring because it just lives inside everything. We measure it, then we replace it.
How the engine works
Any format the work arrives in.
PDFs, scans, photos, emails, spreadsheets, EDI, faxes, the document type your team complains about every Monday. We meet documents where they land, before they become a queue.
Know what it is before deciding what to do with it.
Document-type detection, sub-type routing, priority and urgency signals. The classifier is the first thing we tune for your domain, because every downstream decision rides on it.
Fields, tables, signatures, line items, every cell.
Layout-aware extraction across forms, contracts, and unstructured text. Tables come out as tables, signatures as signatures, dates as ISO. The output is what a downstream system can consume, not a transcript.
Confidence per field, validated against the rules that matter.
Per-field confidence, cross-field consistency, lookups against your master data, and policy checks before anything posts. The fields the model is unsure about go to a human queue, with reasoning attached.
Into the system that needs it, with the audit it requires.
ERPs, CRMs, claims systems, ticketing, mainframes, custom internal APIs. Every routed payload is timestamped, hash-sealed, and traceable to the source page it came from.
A glimpse of the work
Extracted
Illustrative extraction trace · invoice
What you get
Cited
Every value traces to a page and a bounding box
Checked
Validated against your source records before it posts
Routed
Uncertain fields reach the right reviewer, with reasons
Built for every sector
Seven industries, the documents that run them, and what we do with each.
How we work
A short discovery, a measured rollout, and a long, quiet operate.
Discover
We sit with the team that processes the documents today, watch the work, and find where time and accuracy quietly leak.
Train
Domain models fine-tuned on your document mix. Layouts, vocabulary, and edge cases captured before we go anywhere near production.
Deploy
Rolled out gradually, with confidence thresholds set conservatively. We watch the first documents together, not over email.
Operate
Exception queues, model recalibration, regression tests against real traces. The system gets quietly better while the team gets visibly faster.
A note from the practice
The model isn't done when it's accurate. It's done when the field it's unsure about gets the right human, with the right context, before the downstream system needs an answer. That's the part we work on after the easy fields.
The Aelix Echo document practice
When a person shouldn't be typing it in, we're here.
45-minute working session · One real document type, extracted live