Systems Architecture & Pipeline Specs

How PeelitNow Deconstructs the Web

A look inside the multi-tier extraction pipeline that converts bloated, ad-cluttered web pages and complex research papers into high-fidelity executive summaries.

Ingestion Flow Diagram

RFC 1918 / AST / Bedrock Nova

[Client Input] ──► [DNS & SSRF Shield] ──► [AST DOM Pruning] ──► [Readability Scoring]
                          │ (Drop RFC 1918)          │ (Drop <script>, <nav>)     │ (Text/Link Ratio)
                          ▼                          ▼                            ▼
                 [403 Security Halt]       [15MB Payload Guard]           [Clean Markdown]
                                                                                  │
                                                                                  ▼
[DynamoDB Cache] ◄── [Deterministic Hash] ◄── [Bedrock Nova Micro] ◄──────────────┘
 (Sub-50ms TTL)       (SHA-256 Normalized)     (4 Structured Invariants)

Pipeline Stages & Verification Controls

STAGE 01

Target Ingestion & SSRF Defense Shield

Network Security

Every incoming URL passes through synchronous DNS resolution before any network socket is opened. Private subnets (RFC 1918), loopback interfaces (127.0.0.1), link-local endpoints, and cloud instance metadata services (169.254.169.254) are rejected immediately. Up to 5 HTTP redirect hops are monitored with re-validation on every step under a 15MB payload cap.

STAGE 02

AST Parsing & Readability Heuristics

DOM Normalization

Raw HTML responses are parsed through an Abstract Syntax Tree (AST) processor that strips non-editorial nodes, including <script>, <style>, <nav>, <aside>, hidden tracking pixels, and nested ad iframes. Mozilla Readability heuristics evaluate comma frequency, paragraph density, and link-to-text ratios to isolate the core semantic body.

STAGE 03

Map-Reduce PDF & Media Processing

Document Parsing

For multi-page whitepapers and research preprints, client uploads stream directly to private S3 storage via short-lived pre-signed URLs. The serverless PDF processor (unpdf) extracts text per page, chunks sections into 8,000-token windows, and executes a multi-pass Map-Reduce algorithm to assemble dense academic findings without context window truncation.

STAGE 04

Amazon Bedrock Nova Micro Inference

LLM Orchestration

Normalized plain text is structured into token-clamped prompt envelopes dispatched to Amazon Bedrock running the Amazon Nova Micro model via secure IAM container roles. The model is constrained to 4 structural invariants: Delta (what changed), Mechanism (how it works), Metrics (empirical outcomes), and Constraints (limitations).

STAGE 05

Deterministic Caching & Edge Delivery

Data Layer

Generated peels are stored in Amazon DynamoDB using canonical SHA-256 URL hashing. Subsequent requests for the same source document are delivered in sub-50ms directly from cache, reducing Bedrock API overhead and eliminating redundant scraping load on upstream publishers.

Operational Benchmarks & Latency Budgets

Execution Phase	Target Budget	Technology Layer	Safety Threshold
DNS Resolution & SSRF Sweep	< 45 ms	Node.js DNS Async Lookup	RFC 1918 / IMDS Rejection
DOM Parsing & Content Extraction	120 – 250 ms	@mozilla/readability + Cheerio	15 MB Payload Cap
LLM Inference & Synthesis	950 – 1,800 ms	AWS Bedrock Nova Micro	4 Structured Takeaways
Cached Lookup Hit Rate	< 50 ms	Amazon DynamoDB	24h – 30d TTL Eviction

Try the live extraction engine

Submit an article or academic URL to evaluate the summary fidelity.

Peel an Article Now