Skip to content

SYSTEM Cited by 1 source

Doczy.ai

Doczy.ai is AArete's intelligent contract interpretation SaaS solution — an LLM-grounded document-intelligence platform built on AWS that converts unstructured contracts and legal documents into structured, queryable repositories. Healthcare-focused but applied to financial services as well; integrates with existing CLM (Contract Lifecycle Management) systems (Coupa, Icertis) as an upstream extraction substrate rather than as a CLM replacement.

What it is

A production document-intelligence pipeline that has, over 22 months, processed:

  • 2.5 million contract documents (50 million pages)
  • 137 million Amazon Bedrock API calls
  • 442 billion tokens (~3 200 tokens/call average)
  • ~250 000 contracts/week peak production throughput
  • ~$330 million cumulative client savings (direct + indirect)
  • 99% extraction accuracy (vs ~55% for AArete's pre-2024 rules-based predecessor)
  • 97% reduction in manual processing time

(All figures from sources/2026-06-02-aws-automating-contract-intelligence-with-doczyai-on-aws, which is the canonical and only public AArete-published architecture disclosure on the wiki to date.)

Architecture (AWS service composition)

Nine-service AWS pipeline plus Snowflake as analytical sink:

Cognito (authn) → Next.js frontend
                         │
                         ▼
                   S3 (document upload)
                         │  S3 event
                         ▼
                   Lambda (orchestrator)
                         │
                         ▼
                   Textract (OCR + metadata)
                         │
                         ▼
            ┌────────────────────────────┐
            │  Doczy.ai processing layer │
            │  (likely on ECS containers)│
            │                            │
            │  1. Smart chunking         │
            │  2. Dual clustering engine │
            │     (semantic + structural)│
            │  3. File-class detection   │
            │  4. Domain-tuned prompt    │
            └────────────────────────────┘
                         │
                         ▼
                   Bedrock (LLM extraction)
                         │
                         ▼
                   Snowflake (structured repo)
                         │
                         ▼
                Intelligent dashboards
                         │
                         ▼
                Downstream automation
                  - Claims systems
                  - Vendor invoice
                    verification

Cross-cutting:
  CloudWatch (real-time monitoring)
  Secrets Manager (sensitive config)

Canonicalised as patterns/managed-ai-document-intelligence-pipeline-on-aws.

Three load-bearing algorithmic primitives

1. Smart chunking (concepts/smart-chunking)

"a proprietary approach that goes far beyond pulling words off a page. Rather than treating a document as a flat sequence of text, smart chunking preserves hierarchical structure and one-to-many relationships within documents."

Disclosed properties:

  • Hierarchy preservation across chunk boundaries
  • One-to-many relationship preservation (e.g. parent clause → multiple sub-clauses, exhibit → multiple referenced schedules)
  • Semantic + keyword decomposition combined
  • Dynamic parameters to maintain logical relationships
  • Sequential identifiers + metadata-driven grouping organise chunks into field groups
  • Overlap detection + duplication removal

Mechanism details (algorithm class, dynamic-parameter tuning, ID scheme) are AArete IP and not disclosed.

2. Dual clustering engine (concepts/dual-clustering-document-intelligence)

"This two-lens methodology analyzes every contract simultaneously from both a semantic and a structural perspective."

Lens Mechanism What it captures
Semantic Embeddings → similarity grouping Meaning, paraphrases, conceptual equivalence
Structural Pattern-recognition algorithms Clause types, formatting conventions, table layouts, hierarchical depth

The two clusterings run in parallel; projection algorithms compare the semantic and structural clusters side by side, fusing them into a "unified, enriched document model that captures both meaning and context."

The article positions dual clustering as the load-bearing contributor to the 99% accuracy — "It's this convergence that drives the 99% accuracy rate of Doczy.ai. The system doesn't just read the words, it understands the contract." The structural lens catches things the semantic lens misses (a "three-nested-level exhibit carries fundamentally different implications than a straightforward attached schedule"); the semantic lens catches things the structural lens misses (paraphrased clauses with different formatting but the same meaning).

3. Prompt optimisation feedback loop (concepts/prompt-optimization-feedback-loop)

"Before output is finalized, the system determines each document's file class and generates prompts tailored to the extracted text, cluster classification, and domain context. Through few-shot and multi-shot prompting, the platform continuously edits the prompt on domain-specific examples and based on real outputs, creating a feedback loop that compounds accuracy improvements over time."

Two structural sub-primitives:

  • File-class routing — document type detected first; per-class prompt template selected
  • Continuous prompt editing based on production outputs — "compounds accuracy improvements over time"

This is not model fine-tuning — it operates entirely at the prompt layer.

Three-era accuracy progression

Era Approach Accuracy Throughput
Pre-2020 Manual review by individuals not disclosed ~100 docs/week per individual
2020–2023 Rules-based contract processing ~55% not disclosed
2024+ AI-based on AWS (Doczy.ai) 99% ~250 000 docs/week production

The article positions the 55%→99% jump as the structural payoff of replacing rules with LLM-grounded extraction, with smart chunking + dual clustering as the grounding substrate that lets the LLM see the document's structure rather than a flat token sequence.

Downstream consumers

Two named health-plan automation capabilities:

  1. Claims-system configuration — auto-translate reimbursement terms into [claims-system rules]
  2. Vendor invoice verification — assess payments against contract terms; flag overpayments / underpayments before they occur

CLM-system integrations: Coupa, Icertis. Doczy.ai sits upstream of CLM systems as an extraction substrate, not as a CLM replacement.

Delivery model

SaaS — AArete operates the AWS architecture for the customer. "AArete's team of experts will configure this solution for your specific document types, domain terminology, and business processes, supporting maximum value from day one." Customer-side configuration burden = identifying document classes and domain terminology; AWS infrastructure burden = AArete-side.

Caveats

  • Bedrock foundation model identity not disclosed (Claude / Titan / Llama / Mistral; the 3 200-tokens-per-call average is consistent with several SKUs).
  • Smart chunking and dual clustering algorithm internals not disclosed (AArete IP / patent).
  • Multi-tenancy isolation mechanism not disclosed.
  • HIPAA / BAA / PHI handling posture not addressed at architectural altitude.
  • No latency envelope per document.
  • No comparison vs alternative document-intelligence vendors (Icertis Discovery, Kira / Litera, Evisort, etc.).
  • No retention / privacy / encryption posture for processed document data.

See sources/2026-06-02-aws-automating-contract-intelligence-with-doczyai-on-aws for the full caveat enumeration.

Seen in

Last updated · 542 distilled / 1,571 read