Amazon Bedrock

Upwork Job Post: Build “Project Aegis” — Amazon Bedrock AI System for Estate Planning Document Review Job Title AWS Bedrock AI Engineer Needed to Build Legal Drafting Review System for Estate Planning Packets Project Name Project Aegis Overview We are a Florida law firm looking to build an internal AI quality-control system named Project Aegis. The goal of Aegis is to help our firm review estate planning packets before they are finalized or sent to clients. These packets usually include documents such as: * Revocable trusts * Wills * Durable powers of attorney * Health care surrogate designations * HIPAA authorizations * Living wills * Related estate planning documents * Engagement letters The packets are usually 120 to 200 pages long and may be in English only or English and Spanish. Aegis should review the drafted packet against the client’s actual information and communications to catch drafting mistakes, inconsistencies, wrong names, repeated names, cross-client contamination, and mismatches between the source information and the final documents. This is not a generic chatbot. This is a secure legal-document review system that must compare drafted documents against source materials and produce a clear issue report for attorney/staff review. Main Goal Build an internal AI system using Amazon Bedrock that can review estate planning packets and flag possible drafting errors before the documents are finalized. The system should compare the estate planning packet against available source materials, including: * Intake forms * Meeting transcripts * Cloud meeting videos * Zoom phone call recordings/transcripts * Outlook emails * Microsoft Teams meetings * Microsoft Teams chat messages * Prior client communications * Related engagement letters * Firm-approved templates/forms The final output should be a structured quality-control report showing possible errors, the document/page/section where the issue appears, the source that supports the correction, and the suggested fix. Examples of Errors Aegis Should Catch Aegis should identify errors such as: 1. Wrong Person in Wrong Role Example: The wife’s name is placed where the child’s name should appear. The system should detect that the role does not match the client intake, meeting transcript, or other source materials. 2. Inappropriate Repetition of Names Example: The wife is listed correctly as spouse, but her name is also mistakenly repeated as a child, trustee, beneficiary, or other role where she does not belong. 3. Cross-Matter Contamination Example: A name from another client matter appears in the current client’s estate planning packet. This is a major issue. The system should flag names, addresses, relatives, fiduciaries, or other facts that do not belong to the current matter. 4. Incorrect Family Relationships Example: A person identified as the client’s child in the draft is actually the client’s spouse, sibling, parent, or unrelated person according to the intake or meeting. 5. Engagement Letter Mismatches The same types of errors can occur in the engagement letter. Aegis should review the engagement letter for: * Wrong client name * Wrong spouse name * Wrong scope of work * Wrong matter type * Names from another client matter * Wrong address * Inconsistency with the estate planning packet 6. Bilingual Document Inconsistencies Some packets are English only. Others include English and Spanish. Aegis should detect when the English version and Spanish version do not match on key information, including: * Names * Roles * Fiduciaries * Beneficiaries * Property references * Distribution terms * Client instructions * Engagement terms 7. Missing or Placeholder Information The system should flag blanks, placeholders, bracketed text, template remnants, or incomplete fields. Examples: * “[Client Name]” * “[Spouse Name]” * “INSERT ADDRESS” * Wrong template language * Unfilled notary or witness sections * Inconsistent dates 8. Contradictions Between Drafted Documents The system should compare all documents in the packet against each other. Example: * Trust says Child A receives 100% * Will or memo says Child A and Child B share equally * Power of attorney names Wife as agent * Intake says Child is supposed to be agent 9. Contradictions Between Drafts and Client Communications The system should compare drafted documents against intake forms, emails, meeting transcripts, Teams chats, and phone calls. Example: * Client said in the meeting that daughter Maria should be successor trustee * Draft names son Carlos as successor trustee * Engagement letter references a joint trust, but documents are drafted as individual trusts Required Data Sources The system should be designed to ingest, search, and compare against: 1. Estate planning packet documents 2. Engagement letters 3. Intake forms 4. Zoom meeting transcripts and cloud meeting recordings 5. Zoom phone call recordings and transcripts 6. Outlook emails 7. Microsoft Teams meeting transcripts/recordings 8. Microsoft Teams chat messages 9. Firm templates and premade forms 10. Optional: prior versions of the same documents Preferred Technical Architecture We are looking for someone who can recommend and build the right architecture, but we expect the project may involve: * Amazon Bedrock * Amazon Bedrock Knowledge Bases * Amazon Bedrock Agents, if appropriate * Amazon S3 * Amazon Transcribe or equivalent transcription pipeline * Amazon Textract or equivalent OCR/document extraction * Vector database such as OpenSearch Serverless, Aurora pgvector, or another appropriate tool * AWS Lambda or containerized ingestion services * IAM least-privilege permissions * KMS encryption * CloudWatch logging * Microsoft Graph API * Outlook / Microsoft 365 integration * Microsoft Teams integration * Zoom API * Secure document parsing for Word and PDF files We are open to the developer’s recommendations, but the solution must be secure, maintainable, and appropriate for confidential law firm data. Required Features 1. Estate Planning Packet Upload and Review Staff should be able to upload or select an estate planning packet and ask Aegis to review it. The review should include: * Client identity check * Spouse identity check * Children and family member check * Fiduciary role check * Beneficiary check * Distribution check * Address/property check * Document consistency check * Engagement letter consistency check * Bilingual consistency check, when applicable * Cross-client name contamination check * Placeholder/template-remnant check 2. Source-of-Truth Extraction Aegis should extract a “source-of-truth” profile for the matter from the available materials. This should include: * Client name * Spouse name * Children * Other family members * Trustees * Personal representatives * Agents under power of attorney * Health care surrogates * Beneficiaries * Distribution instructions * Property addresses * Special instructions * Scope of engagement * Important meeting statements or client instructions 3. Document Comparison Aegis should compare the source-of-truth profile against the drafted documents and identify discrepancies. 4. Structured Error Report The output should be a structured report with: * Issue number * Severity level: Critical / High / Medium / Low * Document name * Page number or section, if available * Problem found * Why it may be wrong * Correct source information * Source citation * Suggested correction * Whether attorney review is required Example output: Issue 3 — Critical Document: Revocable Trust Section: Article 4, Beneficiaries Problem: The draft lists “Maria Lopez” as the client’s child. Concern: Intake form identifies Maria Lopez as the client’s spouse, not child. Source: Intake form dated [date]; Zoom transcript dated [date]. Suggested correction: Review beneficiary section and confirm intended child/beneficiary. Attorney review required: Yes. 5. No Unsupported Legal Conclusions Aegis should not make final legal judgments. It should identify possible drafting issues for attorney or staff review. The system should use language such as: * “Potential inconsistency” * “Possible drafting error” * “Needs attorney review” * “Source does not support this name/role” * “Insufficient information to confirm” 6. Source Citations Required Every flagged issue should cite the source material used to identify the potential problem. Sources may include: * Intake form * Meeting transcript * Email * Teams chat * Zoom call transcript * Engagement letter * Prior version of document * Firm-approved template 7. Cross-Matter Protection This is very important. Aegis should help detect if a name, address, family member, fiduciary, or other client-specific information from a different matter appears in the current packet. The system should be designed to reduce the risk of false positives but should flag suspicious names that are unsupported by the current matter’s source materials. 8. Confidentiality and Security This is a law firm system. Confidentiality is essential. The solution must include: * Encryption at rest and in transit * Secure AWS architecture * IAM least-privilege access * Matter-level access control * Audit logs * No unnecessary third-party data sharing * No use of client data to train public models * Ability to delete/exclude data * Secure handling of transcripts, recordings, emails, and legal documents * Documentation of all data flows Bonus Feature: Document Drafting Bonus points if the developer can also build a system where Aegis can help draft estate planning documents using our premade forms/templates. This should be treated as a possible Phase 2 or optional feature. The drafting system would ideally: * Use our approved templates/forms * Pull client information from intake forms and source materials * Populate the correct fields * Avoid cross-client contamination * Produce a draft for human review * Generate a change log or source map showing where each inserted fact came from * Never finalize documents without attorney/staff approval We are especially interested in developers who understand the difference between: 1. AI-assisted document drafting, and 2. AI quality-control review of attorney-drafted documents. The first priority is quality-control review. Drafting is a bonus. Deliverables The selected freelancer or agency should deliver: 1. Architecture proposal 2. Security design 3. Data-flow diagram 4. AWS Bedrock environment setup 5. Document ingestion and parsing system 6. Intake/source-of-truth extraction system 7. Estate planning packet review workflow 8. Engagement letter review workflow 9. Bilingual review approach 10. Cross-matter contamination detection 11. Structured issue-report generator 12. Source citation system 13. Test dataset using dummy or sanitized matters 14. Pilot implementation 15. Documentation 16. Staff training/handoff 17. Cost estimate and AWS cost-control recommendations 18. Optional Phase 2 proposal for document drafting from templates Suggested Milestones Milestone 1: Discovery and Architecture * Review our current document workflow * Identify data sources * Confirm AWS architecture * Confirm Microsoft 365, Outlook, Teams, Zoom, and document storage requirements * Design the matter-level security model * Provide implementation plan and estimated AWS cost range Milestone 2: Proof of Concept * Use dummy or sanitized estate planning packets * Ingest sample intake forms, transcripts, emails, and documents * Demonstrate detection of wrong names, repeated names, and cross-matter contamination * Produce a structured review report with citations Milestone 3: Estate Planning Packet Review MVP * Build the main document review workflow * Review 120 to 200 page packets * Extract key names, roles, fiduciaries, beneficiaries, and distribution terms * Compare drafted documents against source materials * Generate issue reports Milestone 4: Engagement Letter Review * Add review of engagement letters * Compare engagement letters against intake, client communications, and estate planning packet * Flag wrong names, wrong matter scope, wrong parties, and cross-client contamination Milestone 5: Communications and Meeting Data Integration * Add ingestion from Outlook emails * Add ingestion from Microsoft Teams meetings and chats * Add ingestion from Zoom meeting transcripts and Zoom phone calls * Match communications to the correct matter Milestone 6: Security, Testing, and Pilot Rollout * Implement permissions * Implement audit logs * Test with a limited number of real or sanitized matters * Tune issue severity levels * Reduce false positives * Improve report quality Milestone 7: Production Rollout and Handoff * Deploy production version * Train staff * Provide documentation * Provide maintenance plan * Recommend future improvements Optional Milestone 8: AI-Assisted Drafting From Firm Templates * Analyze firm templates * Build controlled document-generation workflow * Populate templates using source materials * Generate drafts for attorney/staff review * Produce source map for inserted facts * Add safeguards against cross-client contamination Required Experience Please apply only if you have experience with several of the following: * Amazon Bedrock * Amazon Bedrock Knowledge Bases * Retrieval-Augmented Generation * Legal document AI review * Long-document review * AWS S3 * AWS IAM * AWS KMS * AWS Lambda * Vector databases * Amazon Transcribe or audio transcription * Amazon Textract or OCR/document extraction * Microsoft Graph API * Outlook / Microsoft 365 integration * Microsoft Teams integration * Zoom API * Word and PDF document parsing * Bilingual English/Spanish document workflows * Secure enterprise AI systems * Confidential-data environments such as legal, healthcare, finance, or insurance Strong Preference We prefer someone who has already built: * AI document review systems * RAG systems with citations * AWS Bedrock applications * Legal technology tools * Microsoft 365 integrations * Systems that compare drafts against source materials * Systems that detect inconsistencies across long documents * Secure professional-services knowledge systems Important Rules * Do not propose a generic ChatGPT wrapper. * Do not propose uploading confidential client data to random third-party AI tools. * Do not propose fine-tuning on client data unless you can clearly explain why it is necessary. We believe a RAG and source-comparison approach is likely better. * Do not apply if you cannot explain security and permissions. * Do not apply if you cannot build source-cited issue reports. * Do not apply if you cannot work with long legal documents. * Do not apply with a generic AI-generated cover letter. * Do not ask for confidential client data during the proposal stage. Initial demos must use dummy or sanitized data. Screening Questions Please answer the following in your application: 1. Have you built an Amazon Bedrock RAG or document-review system before? Please describe it. 2. Have you built an AI system that reviews long documents for inconsistencies? 3. How would you review a 120 to 200 page estate planning packet? 4. How would you extract the correct client names, family members, fiduciaries, beneficiaries, and roles? 5. How would you detect that the wife’s name was mistakenly placed where a child’s name should be? 6. How would you detect repeated names in inappropriate roles? 7. How would you detect names from another client matter appearing in the wrong file? 8. How would you compare the drafted packet against intake forms, transcripts, emails, and Teams messages? 9. How would you handle English/Spanish bilingual packets? 10. How would you make the AI cite its sources for every issue? 11. How would you reduce hallucinations and false positives? 12. What AWS services would you use and why? 13. How would you handle confidentiality and matter-level permissions? 14. Have you worked with Microsoft Graph, Outlook, Teams, or Zoom APIs? 15. Can you also build a controlled document-drafting workflow using premade templates? If yes, describe your approach. 16. What are the biggest risks or limitations you see in this project? Application Instruction To show that you read this post, start your proposal with this sentence: “I understand that Project Aegis is a legal drafting quality-control system, not a generic chatbot.” Then provide: * Your recommended architecture * Similar projects you have built * Your proposed milestones * Estimated timeline * Estimated budget or pricing structure * Questions you need answered before starting Confidentiality Note We are a law firm. Do not request confidential client information during the proposal stage. We will only use dummy, sample, or sanitized data until proper confidentiality protections, access controls, and project agreements are in place.