AWS Builder Center
Building The Agentic Research Reviewer: An Adversarial Multi-Agent Swarm on Amazon Bedrock and AWS Fargate

Building The Agentic Research Reviewer: An Adversarial Multi-Agent Swarm on Amazon Bedrock and AWS Fargate

Discover how The Agentic Research Reviewer uses an autonomous five-agent swarm on Amazon Bedrock, AWS Textract, and DynamoDB to stress-test scientific claims, detect hidden methodological flaws, and audit reproducibility in academic preprints.

1. The Crisis in Scientific Literature: Why Passive AI Summarization Fails

Over five million academic papers and preprints are published every year across arXiv, bioRxiv, PubMed, and institutional archives. For researchers, doctoral candidates, and machine learning engineers, keeping pace with this literature is an overwhelming operational challenge.
When researchers turn to standard commercial AI assistants for help, they encounter a fundamental architectural flaw: commercial chatbots are tuned to be polite, passive summarizers. They accept empirical benchmarks at face value, uncritically regurgitate author assertions, and overlook critical methodological omissions:
  • Unreleased Code Repositories: Papers claiming breakthrough performance without releasing model weights or training pipelines.
  • Statistically Underpowered Baselines: Evaluating algorithms across small sample sizes (N=6, p > 0.05) without adequate statistical power.
  • Compute Disparities: Comparing fine-tuned proprietary architectures against undertrained public baselines with mismatched compute budgets.
  • Hidden Failure Points in Appendices: Breakthrough throughput or speedup claims that quietly collapse under Out-of-Memory (OOM) errors once sequence lengths exceed trivial limits (N > 64).
When researchers attempt to replicate these papers in production or physical labs, they frequently waste months discovering that the published claims do not hold.
To solve this reproducibility crisis, I engineered The Agentic Research Reviewer: an autonomous adversarial peer-review and claim-verification engine on Amazon Web Services that rigorously interrogates research papers before valuable lab time and compute budgets are spent.



2. High-Level Architecture: Decoupled Multi-Agent Orchestration

The application is architected as an event-driven, decoupled system hosted on an Amazon Elastic Container Service (Amazon ECS) cluster with AWS Fargate, orchestrating an autonomous swarm on Amazon Bedrock, AWS Textract, Amazon S3, and Amazon DynamoDB.
[ React 19 Client Dashboard ]
|
| (PDF Upload via S3 Presigned URL)
v
[ Amazon S3 Bucket: paper-reviewer-storage ]
|
v
[ AWS Textract: Layout & Tables Parsing ]
|
v
[ 3-Page Windowed IDP Engine: Context Chaining ]
|
v
[ Amazon Bedrock Multi-Agent Swarm (Converse API) ]
|--\> Agent 1: Structural Extractor (Amazon Nova Micro)
|--\> Agent 2: Adversarial Critic (Anthropic Claude Opus 4.5)
|--\> Agent 3: Evidence Verifier (Meta Llama 3.3 70B)
|--\> Agent 4: Concept Visualizer (Amazon Nova Lite)
|--\> Agent 5: Adaptive Communicator (Amazon Nova Pro)
|
v
[ Amazon DynamoDB: Atomic Session Checkpoints (\<10ms) ]
|
v
[ Hydrate React 19 Dashboard & Claim Decision Matrix ]



3. Ingestion Engineering: Defeating the Multimodal Token Trap with AWS Textract

Scientific manuscripts present complex two-column layouts, mathematical equations, embedded charts, and numerical pipe tables.
Traditional approaches feed raw PDF page screenshots into multimodal vision models. For a 24-page paper, this burns over 14,000 visual tokens costing $1.85 to $2.50 per paper, while frequently scrambling column order and dropping table headers.
To eliminate this token bloat, I engineered a 3-Page Windowed Intelligent Document Processing (IDP) Engine:
  1. AWS Textract Layout and Tables API: Linearizes multi-column text into standardized Markdown and extracts tabular structures at $0.0015 per page.
  1. Forward-Chaining Context Windows: Processes the manuscript in rolling three-page windows where each chunk carries a cumulative summary of prior sections. This preserves mathematical theorem continuity and equation definitions across extended appendices.
  1. 95% Ingestion Cost Reduction: Drops document ingestion costs from $1.85 down to $0.058 per paper, achieving an 80% token deflation while maintaining complete structural precision.
export async function linearizeDocumentWithContext(pdfBuffer: Buffer): Promise\<LinearizedManuscript\> {
const textractBlocks = await textractClient.send(new AnalyzeDocumentCommand({
Document: { Bytes: pdfBuffer },
FeatureTypes: ["TABLES", "LAYOUT"]
}));
const pages = parseTextractLayout(textractBlocks);
let runningContext = "";
const structuredSections = [];
for (let i = 0; i \< pages.length; i += 3) {
const windowSlice = pages.slice(i, i + 3);
const windowMarkdown = formatToMarkdown(windowSlice, runningContext);
runningContext = await updateRunningSummary(runningContext, windowMarkdown);
structuredSections.push(windowMarkdown);
}
return { fullText: structuredSections.join("\n\n"), pageCount: pages.length };
}



4. The 5-Agent Swarm on Amazon Bedrock: Specialized Multi-Model Routing

Rather than relying on a single monolithic model, The Agentic Research Reviewer dispatches tasks across a heterogeneous foundation model swarm on Amazon Bedrock using the unified Converse API:
  • Agent 1: Structural Extractor (Amazon Nova Micro): Sub-second extraction of falsifiable empirical claims, baseline constraints, and dataset declarations. Blazing speed and deterministic JSON adherence.
  • Agent 2: Adversarial Critic (Anthropic Claude Opus 4.5): Industry-leading benchmark performance on scientific counterfactual deduction, identifying unreleased code, baseline compute mismatches, and overclaimed generalizability.
  • Agent 3: Evidence Verifier (Meta Llama 3.3 70B): Audits mathematical theorem proofs, calculates statistical power (p < 0.05), and constructs operational IF-THEN boundary limit cards.
  • Agent 4: Concept Visualizer (Amazon Nova Lite): Generates interactive system architecture flowcharts and four-pillar mindmaps mapping token transformations through attention kernels.
  • Agent 5: Adaptive Communicator (Amazon Nova Pro): Synthesizes complex mathematical findings across eight specialized academic personas (PhD Researcher, ML Systems Engineer, Statistician, Grant Reviewer, Industry Practitioner, Undergrad, Peer Reviewer, and General Scientist).

Why Amazon Bedrock?

  1. Unified Converse API Perimeter: Eliminates fragmented third-party SDK wrappers. A single API interface manages model routing, request schemas, and stop reasons across Amazon Nova, Anthropic, and Meta models.
  1. Zero Customer Data Training: Proprietary drug candidate formulas, patent-pending algorithms, and unpublished preprints never leave customer cloud boundaries or get used for model pre-training.
  1. Cross-Region Inference Profiles: Guarantees automated resilience and seamless failover routing during peak submission traffic.



5. Resilient State Persistence with Amazon DynamoDB and S3

A common failure in multi-agent workflows is the cascading failure trap: if an agent times out, all earlier agent work is lost.
To guarantee fault tolerance:
  • Amazon DynamoDB (PaperReviewerSessions): Stores atomic state checkpoints after every agent execution with sub-10ms read/write latency. If a connection drops, the review resumes from the last completed agent step without re-invoking previous models.
  • Amazon S3 (paper-reviewer-storage): Provides high durability for raw PDF manuscripts, linearized Markdown documents, and cropped figure assets. S3 presigned URLs allow secure direct-to-bucket client uploads without passing multi-megabyte payloads through API servers.



6. Containerized Deployment: Amazon ECS with AWS Fargate

The application runs as a containerized microservices stack deployed on Amazon ECS backed by AWS Fargate:
version: '3.8'
services:
agentic-reviewer:
image: ${AWS_ACCOUNT_ID}.dkr.ecr.${AWS_REGION}.amazonaws.com/paper-reviewer:latest
cpu: 2048
memory: 4096
environment:
- AWS_REGION=${AWS_REGION}
- S3_BUCKET_NAME=paper-reviewer-storage
- DYNAMODB_TABLE_NAME=PaperReviewerSessions
ports:
- "3000:3000"
healthcheck:
test: ["CMD", "curl", "-f", "http://localhost:3000/api/health"]
interval: 30s
timeout: 5s
retries: 3
  • Application Load Balancer (ALB): Routes traffic across auto-scaled Fargate tasks with SSL certificates managed via AWS Certificate Manager (ACM).
  • IAM Role Isolation: Each Fargate task assumes an ephemeral IAM role granting least-privilege permissions (bedrock:InvokeModel, textract:AnalyzeDocument, dynamodb:UpdateItem, s3:PutObject). No static API keys are ever stored in code.



7. What Researchers Experience: The Review Dashboard

When a manuscript is analyzed, researchers interact with five core verification modules:
  1. Claim vs. Evidence Matrix: Deconstructs claims into green and red IF-THEN Boundary Limit Cards (e.g., HOLDS when sequence length N <= 64; COLLAPSES under Out-Of-Memory when N > 512 due to quadratic memory explosion).
  1. Critical Methodological Question Inventory: Provides 5+ line analytical citations with mathematical notation (O(N^2), p > 0.05) interrogating baseline parity and sample size power.
  1. Concept and Evidence Map: Interactive vector flowcharts with 50% to 250% zoom controls illustrating the authentic data pipeline.
  1. Data Availability and Missing Artifact Auditor: Automatically inspects GitHub links, model weight repositories, and supplemental appendices to flag missing assets.
  1. Original PDF Document Viewer: An embedded viewer modal displaying structural metadata metrics (Pages, Figures, Tables, Citations, Verified Claims) alongside the manuscript.



8. Feedback and What Could Be Improved in AWS Services

  1. Amazon Bedrock Converse API JSON Schema Enforcement: While the Converse API successfully unifies heterogeneous model providers, output formatting can still exhibit subtle schema variations across model families. Introducing native, guaranteed JSON schema validation directly within the Converse API runtime across all foundation models would simplify multi-agent tool calling pipelines.
  1. AWS Textract Mathematical Notation Recognition: AWS Textract Layout and Tables is unmatched at multi-column academic text and tabular parsing. However, complex inline math symbols (subscripts, tensor indices, nested integrals) are occasionally simplified into unicode text. Adding native LaTeX math block recognition would establish Textract as the definitive tool for scientific document analysis.
  1. Streamlined IAM Templates for ECS Cross-Region Inference: Configuring IAM task execution policies for cross-region inference profile ARNs (e.g., us.amazon.nova-pro-v1:0) involves navigating multiple documentation pages. Providing pre-built IAM policy templates directly in the Amazon ECS console for Bedrock multi-agent setups would accelerate development.



9. Conclusion

Modern scientific progress moves too fast to waste valuable researcher hours and compute budgets on irreproducible or overclaimed preprints. By pairing AWS Textract's structural layout parsing with an autonomous 5-agent swarm on Amazon Bedrock, The Agentic Research Reviewer replaces polite chatbot flattery with rigorous, mathematical, and adversarial scientific auditability.
All code, API specifications, and automated test suites are open-source and available on GitHub:  github.com/niloydebbarma-code/paper-understanding-helper 
Built by Niloy Deb Barma for the AWS Global Hackathon.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article