ARC Labs · Reference Architectures
ARC Labs 0.2.0 · updated 2026-08-11
DocsRoadmapGitHub
Free · Open repository

Battle-tested system blueprints you can adapt instead of starting from a blank diagram.

A free, open library of reference architectures harvested from ARC's real work — AWS streaming labs, Operator AI, Labs products, and field systems. Each entry is a documented blueprint: the problem, the component shape, key decisions and trade-offs, failure modes, when not to use it, and a scaling path. No account, no lock-in.

20 architectures
Northline Rides — Black Friday surge. Partition by driver_id, Enhanced Fan-Out for fraud, Firehose for the lake.
Dataintermediate
Kinesis Data Streams — Distributed Log
AWS's managed distributed log: shards, partition keys, classic vs Enhanced Fan-Out consumers, and checkpointing. The foundation pattern for real-time pipelines.
#kinesis#streaming#shards#efo
Launch-week ingest: replay plus three consumers picks Kinesis, not SQS.
Datastarter
Streaming Service Decision Tree
When to pick Kinesis Data Streams, Firehose, MSK, SQS, or DynamoDB Streams — decided by consumer count, replay, latency, retention, and toolchain.
#kinesis#msk#sqs#firehose
Launch-day clickstream. Commit the offset after the upsert; poison messages go to the DLQ.
Dataintermediate
Event-Driven Ingestion Pipeline
Durable, replayable ingestion using a queue or log, idempotent consumers, and a dead-letter path for poison messages.
#events#queue#ingestion#idempotency
Avery Consulting POC daily brief. Explicit stages, no multi-agent orchestration.
AIintermediate
Deterministic Operator Pipeline
Explicit ingest → extract → validate → persist → retrieve → brief stages. No multi-agent orchestration in the proof of concept.
#operator-ai#pipeline#bedrock#adr
Meridian Health: reported green, evidence red. The gap is the product.
AIadvanced
Executive Operating Intelligence Loop
Turn fragmented operational evidence into a structured operating model, then render dashboards, briefs, SOWs, and assessments from that model — not from a chatbot.
#operator-ai#intelligence#schema#evidence
S3 is the artifact system of record. The Knowledge Base is a derived index.
AIintermediate
Bedrock Knowledge Base Extraction Pipeline
S3 artifacts into a Bedrock Knowledge Base, schema-validated extraction into DynamoDB, briefs generated on a schedule — the AWS shape behind Operator AI.
#bedrock#rag#s3#dynamodb
Prompt v14 at 2am. Gateway, versioned prompts, traces, and a cost cap.
AIintermediate
LLM Application Backend
A production LLM backend: gateway, prompt/version control, caching, guardrails, and observability.
#llm#ai#gateway#observability
Draft the Jira ticket. A human still has to send it.
AIintermediate
Human-in-the-Loop Action Layer
Draft tickets, emails, and escalations from operating intelligence — but never mutate an external system without a human approval.
#hitl#workflows#approvals#operator-ai
Meridian Health 56-day calendar. Ground truth never enters the model context.
AIadvanced
Synthetic Data and Evaluation Loop
Fictional companies, daily scenarios, and hidden ground truth so you can score extraction quality before any real client data is in the system.
#evaluation#synthetic-data#ground-truth#operator-ai
Fourth free tool, same Next.js + Vercel stack. Three homes: Drive, GitHub, website.
Webstarter
ARC Labs Product Stack
The shared Next.js + Vercel + TypeScript shape every ARC Labs tool ships on, so a new scanner or catalog lands fast and behaves the same.
#nextjs#vercel#labs#platform
Sale-day ticketing homepage. CDN and Redis take the reads; replicas lag.
Webstarter
Read-Heavy Web Service
A cache-fronted, horizontally scaled service pattern for read-dominant workloads with graceful cache invalidation.
#caching#scaling#web#cdn
Franchise CRO crawl of 40 URLs. The web tier returns; workers run the journeys.
Backendstarter
Async Job Processing
Offload slow work to a worker pool with retries, visibility timeouts, and a dead-letter queue.
#jobs#workers#queue#retries
Two agency clients, one database. RLS makes tenant_id mandatory, including on jobs.
SaaSadvanced
Multi-Tenant SaaS Data Isolation
Patterns for isolating tenant data — shared schema with row-level security, up to database-per-tenant.
#multi-tenancy#saas#isolation#security
jordan@ spiked 7.8×. Kill switch via admin APIs — minutes-scale, no proxy.
SaaSintermediate
Agent Spend Control Without a Gateway
Detect-and-cut AI coding-agent spend from read-only admin APIs: per-developer burn, a minutes-scale kill switch, and client chargeback — no traffic proxy.
#finops#agents#kill-switch#tokenloop
Mean looks healthy. Security is 32. A high average with high variance is fragile.
SaaSintermediate
Maturity Assessment with a Balance Score
A 40-question, behaviorally-anchored maturity model that scores pillars *and* variance — a high average with high spread is fragile, not healthy.
#assessment#maturity#cith#scoring
RFP-grade checkout assessment. AI only ranks findings the crawlers produced.
SaaSadvanced
Experience Intelligence Assessment Pipeline
RFP-grade CRO/UX assessment: crawl, Playwright journeys, Lighthouse, axe-core, and AI interpretation into an evidence-backed, prioritized backlog.
#cro#ux#playwright#lighthouse
Charity pickup of 42.5 kg mints Gold-tier points on Postgres — not a chain.
SaaSstarter
Verified Recovery Loyalty Ledger
Verified food recovery mints loyalty points on a Postgres ledger — not a blockchain — with role-specific views for charity, farmer, and impact.
#ledger#loyalty#supabase#foodmesh
Customer endpoint down three hours. Signed retries, then replay.
Integrationintermediate
Reliable Webhook Delivery
Deliver outbound webhooks with retries, signatures, ordering guarantees, and a consumer-visible delivery log.
#webhooks#integration#retries#signing
Server looked fine in one client. Handshake and schemas say otherwise.
Integrationstarter
MCP Conformance Gate
Treat Model Context Protocol servers as production dependencies: handshake, capability negotiation, and schema checks as a gate — not a vibe check in one client.
#mcp#conformance#agents#ci
Prod checkout workload. Six pillars, named owners, a date on the calendar.
Cloudstarter
Well-Architected Review Loop
Run AWS Well-Architected as a repeating operating loop across six pillars — not a one-time slide — with named risks, owners, and a next-workload date.
#aws#well-architected#governance#review
Get notified when new architectures land (optional)

Everything here is free and browsable without an account. This is just for updates.