Return
LegalTalent - Success Story
LegalTalent
From manual to instantaneous: Legal AI in production
LegalTalent transformed its legal compliance platform's document review with a Generative AI solution on AWS, along with Nublit, reducing query time from minutes to seconds with full traceability and control over data.
Collaboration
About the Client
LegalTalent S.A. is a company in the legal sector in Uruguay that manages large volumes of legal documents — contracts, confidentiality agreements, compliance files, and labor clauses — that historically required intensive manual review by lawyers and paralegals. Nublit designed and deployed a production-grade Generative AI platform on AWS that ingests, processes, and indexes legal documents, allowing lawyers to query the entire repository in natural language and receive contextually accurate, cited answers in seconds.
http://compliance.legaltalent.ai Uruguay
Needs
1
Reduce document review time. Manual review took up between 60% and 80% of attorneys' time per case, with the risk of clauses being overlooked and processes that were not scalable as the volume of documents grew.
2
Regain data sovereignty. Reliance on third-party LLM providers raised concerns about data sovereignty, unpredictable costs, and lack of control over model behavior for sensitive legal content.
3
Ensure traceability and regulatory compliance. The manual process left no formal audit trail of which clauses were reviewed or how, creating compliance risks in a highly regulated sector.
What was done?
RAG Platform on Amazon Bedrock
A Retrieval-Augmented Generation architecture was implemented with three Bedrock models in production: Claude Haiku for classification and entity extraction during ingestion, Claude Sonnet for standard legal queries (~85% of volume), and Claude Opus for complex multi-document reasoning.
Automated ingestion pipeline
Documents are uploaded to S3, processed by OCR with Amazon Textract, entity extraction with Bedrock Claude Haiku, and embedding generation with Cohere embed-multilingual-v3 and Amazon Titan — all indexed in Aurora PostgreSQL with pgvector.
Hybrid search
The system combines BM25 (lexical precision for legal terms) with semantic vector search, weighting 0.4/0.6 to maximize accuracy in the legal domain.
End-to-end security and compliance
AWS WAF, KMS with automatic key rotation, GuardDuty Malware Protection on every uploaded document, and 7-year WORM retention on S3 Object Lock for the legal audit trail.
Continuous quality evaluation with RAGAS
Each response is measured against a production dataset of 50 question-answer pairs validated with the client's legal team, with minimum thresholds for Faithfulness, Context Recall, and Answer Correctness before any promotion to production.
Hired tools
Amazon Bedrock (Claude Haiku / Sonnet / Opus) · Amazon Textract · Amazon Aurora PostgreSQL + pgvector · AWS Lambda · Amazon API Gateway · AWS WAF · Amazon GuardDuty · AWS KMS · Amazon EventBridge
Hired tools
Amazon Bedrock (Claude Haiku / Sonnet / Opus)
Amazon Textract
Amazon Aurora PostgreSQL + pgvector
AWS Lambda
Amazon API Gateway
AWS WAF
Amazon GuardDuty
AWS KMS
Amazon EventBridge
Services contracted
Implementation
Cloud Security
Data Architecture
Results
Query speed
The time it took to find a document went from 20–40 minutes of manual searching to less than 3 seconds per query—a reduction of more than 99%.
Legal review time
Reduction of ~60% in total document review time: attorneys consult the platform instead of manually searching through files.
Validated AI Quality
Faithfulness of 0.87 above a threshold of 0.85 — the critical metric for preventing hallucinations in a legal context — along with Context Recall of 0.84 and Answer Correctness of 0.81, evaluated on a real dataset of 50 questions with the client's legal team.
Uncompromising security
0 critical or high findings in AWS Security Hub post-deployment, 100% documents scanned by GuardDuty, and 7-year WORM retention for regulatory compliance
Scale in production
Over 50 legal documents indexed in production since go-live, with the platform active and verified by Amazon Bedrock invocation metrics.
