SYSTEM DESIGN

Topology &
Data Flow.

SortMail separates data ingestion from AI inference. By serving directly from local storage and pushing events over Server-Sent Events, the interface maintains sub-50ms reads regardless of upstream mail provider load.

Push, Don't Poll Data PipelineMermaid Topology
01 / DATABASE-FIRST

Immediate Page Renders

SortMail eliminates blocking third-party API calls during inbox renders. Next.js reads directly from PostgreSQL in under 50 milliseconds while a background worker syncs delta state via Gmail historyId tokens.

02 / SSE STREAMING

Event-Driven Reactivity

Rather than polling the server every few seconds, the client maintains a persistent Server-Sent Events stream. Redis Pub/Sub detects background extraction completions and pushes targeted cache invalidation signals.

03 / SINGLE-PASS LLM

Strict Cost Predictability

Instead of multi-hop autonomous agent loops, SortMail dispatches one dense JSON schema query per thread to Gemini 2.0 Flash, achieving deterministic extraction at approximately ₹0.25 per message.

AI Pipeline Benchmark & Extraction Evidence
Verified Pipeline Spec
Local Read Latency<50ms (PostgreSQL)
Inference Round-Trip<850ms (Gemini 2.0)
Sync StrategyGmail historyId Delta
Security GateStrict MIME Verification
Live Session Trace
$ sortmail process --msg "NDA_Q3.eml"
[info] Fetching thread via historyId: 18F4B2
[info] Cache hit: PostgreSQL local replica ready (38ms)
[info] Invoking Gemini 2.0 Flash single-pass extractor
[✓] Extracted bluf: "Priya Mehta shared updated Q3 agreement for review."
[✓] Deadline Extracted: 2026-10-17T17:00:00Z (Priority: High)
[✓] Extracted action items: 3 tasks routed to tasks table
[✓] Attachment verified: Q3_Analytics_Agreement.pdf (MIME: application/pdf)
[!] Threat check complete: 0 warnings
[info] Published intel_ready on Redis channel: thread_18F4B2 (62ms)