Workflow blueprints

System Design

Classic designs asked in system design interviews, from a URL shortener to ride matching.

• Food delivery: order to doorstepOne order crosses payments, the restaurant, rider dispatch and live tracking, and must never be lost, charged twice or left without a rider.Food delivery
• Cab booking end to end (Uber & Ola)A ride-hailing flow that takes a rider from pickup and drop entry through surge-priced quote, geo-cell driver matching, OTP-verified pickup, live tracking, fare settlement over UPI, cards, wallet or cash, ratings and safety. It must never assign one driver to two trips, charge a fare different from the agreed quote rules, or lose an SOS event.Ride hailing
• API rate limiterA token bucket in front of the API keeps one noisy client from starving everyone else.Public APIs
• Cache-aside cachingThe most common caching pattern: read through the cache, write to the database, then invalidate.E-Commerce
• Event-driven order processingA message broker decouples services so each one scales, fails and retries on its own.E-Commerce
• Scalable web architectureThe classic path of a request, from DNS to replicated data, and where each layer adds scale.Tech & SaaS
• Database sharding & replicationSplitting one huge table across machines by a shard key, with a replica behind every shard.Social & Fintech
• URL shortenerHow a link like sho.rt/abc123 is created once and resolved millions of times.Marketing & SaaS
• Payment processing that never charges twiceIdempotency keys, a double-entry ledger and reconciliation keep money correct when networks fail.Fintech & E-Commerce
• Notification service at scaleOne service turns product events into push, email and SMS, honouring preferences and priorities.Tech & SaaS
• Social news feed with fan-outPrecomputing each follower's feed makes opening the app one fast read, even with millions of users.Social media
• Real-time chat over WebSocketsPersistent connections, ordered messages and presence: the core of every messaging app.Messaging & Collaboration
• Video upload & adaptive streamingUpload once, transcode in parallel, stream at the quality each viewer's network can take.Media & OTT
• Search autocomplete (typeahead)Suggestions on every keystroke, served from a prefix trie rebuilt from real searches.E-Commerce
• Distributed job schedulerCron at scale: due jobs are claimed by one leader, queued by priority and run by idempotent workers.Fintech & Banking
• Ride-hailing driver matchingLive driver positions in a geo index, candidates ranked by road ETA, and an offer that times out.Mobility & Delivery
• E-commerce platform: buy, sell, deliver, returnAn end-to-end marketplace that onboards sellers, sells their stock to buyers, splits orders by seller and warehouse, ships through courier partners, and runs returns, refunds and seller settlement. It must never oversell stock, charge a buyer twice, lose an order after payment, or pay a seller for an item that was refunded.E-Commerce
• School & university management systemA multi-tenant platform that runs admissions, student and staff records, timetables, attendance, fees, courses, exams and portals for schools and universities. It must never leak a student's data to the wrong person, lose or double-count a fee payment, or publish an incorrect or unapproved result.Education
• Campus placement workflowA placement portal that takes students from verified profiles through drives, assessments, interviews and offers, then reports placement statistics. It must never let a student hold offers in breach of the one-offer policy, and it must never make an eligibility or selection decision that cannot be explained from the audit trail.Education
• Movie & event ticket booking (BookMyShow)Lets users pick seats, hold them briefly, pay and receive a QR e-ticket, while a waiting room shields the system during blockbuster openings. It must never sell the same seat for the same show twice, even under a massive rush.Entertainment ticketing
• Hotel reservation & room inventoryGuests search by city and dates, hold specific room types for every night of the stay, pay a deposit and get a confirmed booking, while inventory stays in sync with OTAs. The system must never sell a room-night beyond the physical count plus the explicit overbooking allowance, and must never confirm a booking without inventory decremented for every night.Travel & Hospitality
• Stock trading & order matching engineTakes orders from a broker app through margin and risk checks into a per-symbol order book that matches by strict price-time priority, then publishes confirmations, market data and settlement files. It must never break ordering or fairness, lose an acknowledged order or trade, or let an order through without the margin to cover it.Capital markets
• UPI payment flow end to endMoves money from a payer's bank account to a payee's account through PSP banks and the NPCI UPI switch in seconds. It must never debit twice for one intent, and never leave a debited-but-uncredited payment unresolved.Fintech / Payments
• Distributed cache clusterA Redis-like sharded cache that routes keys with consistent hashing, replicates each shard, evicts by LRU/LFU and TTL, and absorbs hot keys and stampedes. It must never serve data past its intended staleness, and must never let a topology change or mass expiry push full traffic onto the database.Cloud infrastructure
• Cloud file storage & sync (Google Drive, Dropbox)Clients split files into content-hashed chunks, upload only missing chunks to object storage, commit a new file version in a metadata service, and other devices pull changes from a per-user change journal. It must never lose a committed version, silently overwrite a concurrent edit, or expose a file to someone without permission.Cloud storage / SaaS
• Web crawler at scaleDiscovers, fetches, deduplicates and stores billions of web pages, then recrawls them as they change. It must never overload a host, ignore robots.txt, or loop forever in crawler traps.Search and data infrastructure
• Collaborative document editing (Google Docs)Lets many users edit one document at the same time, online or offline, with live cursors, comments, permissions and full version history. It must never lose an acknowledged edit, and every replica must converge to the same document.Productivity software
• Metrics, logs & alerting platformCollects metrics and logs from every service, stores them in tiered time-series and log stores, and turns alert rules into deduplicated pages for the right on-call engineer. It must never silently drop the data or alerts that tell you production is broken, even when the platform itself is under load.Cloud infrastructure
• API gateway with canary releasesA single gateway entry point terminates TLS, authenticates, rate-limits, routes and shields microservices, while shifting a controlled percentage of traffic to new versions. It must never let an unauthenticated request through, and it must never let a bad canary keep receiving traffic after its error budget is breached.Cloud platforms
• 10-minute grocery delivery (dark stores)A quick-commerce workflow that checks whether a location is serviceable, routes the order to one nearby dark store, reserves stock in that store, then runs picking, packing, rider dispatch and live tracking within a 10-minute promise. It must never sell stock the assigned store does not have, and it must never charge for items that were not delivered.Quick commerce
• Real-time gaming leaderboardIngests validated score updates from millions of players into sharded Redis sorted sets. It serves global, regional and friends rank queries in milliseconds and rotates seasons into an archive. It must never show a cheated score or lose a legitimate one, and it must keep ranks consistent while a season resets.Gaming

LangChain

Retrieval, tool-calling agents and extraction pipelines built with LangChain.

• Retrieval-augmented generation (RAG)Index your documents once, then answer each question from the chunks that match it.Corporate knowledge
• Tool-calling agent loopAn agent built with create_agent reasons, calls a tool, reads the result and repeats until it can answer.Tech & SaaS
• Fraud investigation copilotA LangChain agent with read-only tools gathers the evidence, so analysts only make the decision.Banking
• Contract review assistantClause-by-clause review against the company playbook, with redlines a lawyer approves.Legal & Corporate
• Invoice data extractionA Pydantic schema and structured output turn PDFs and scans into checked, typed invoice records.Accounting & Finance
• Text-to-SQL analytics assistantAn agent writes SQL from a plain-English question, behind read-only guardrails, and shows its query.Retail analytics
• Resume screening with structured outputA LangChain create_agent worker redacts personal details with PIIMiddleware, extracts each resume into a validated schema via response_format, and stores evidence-backed scores against a job description for recruiter review. It must never send raw contact details to the model, and it must never let an unvalidated or unexplained score reach a shortlist.HR Tech / Recruitment
• Hybrid search RAG with reranking and citationsAnswers questions about product manuals by running keyword and vector retrieval in parallel, fusing and reranking the results, and generating an answer that cites the exact chunks used. It must never state something the retrieved manual text does not support, and it must refuse when nothing relevant is found.Consumer electronics support
• Customer chat with short- and long-term memoryA banking FAQ assistant built with create_agent that remembers the current conversation through a thread-scoped checkpointer, compresses long chats with SummarizationMiddleware, and recalls customer preferences across conversations from a LangGraph store. It must never leak one customer's memory into another's thread, and must never persist account numbers or other sensitive data as long-term memory.Banking
• Email agent with middleware guardrailsA LangChain v1 email assistant built with create_agent that reads, searches and drafts mail freely, while every send_email call pauses for a human to approve, edit or reject it. PII, summarization and model call limit middleware sit around the model loop. It must never send an email the user did not approve, and must never leak raw PII to the model or the logs.Productivity software
• Agent with tools from MCP serversAn internal operations agent loads ticketing, docs and database tools from several MCP servers through the LangChain MCP adapters, passes them to create_agent, and runs them with per-user authentication and controlled error handling. It must never run a tool with credentials broader than the requesting user's, and it must never perform an unapproved write.Internal IT operations
• Resilient agent with model fallbacks & retriesA customer-facing LangChain v1 agent that keeps answering through provider outages by layering timeouts, model fallback middleware across providers, tool retry, call limits and a cheap-model router for simple turns. It must never repeat a side-effecting tool call such as a refund, and must never leave a customer waiting on a hung request.Customer support
• Web research agent that writes a reportAn analyst agent plans searches, calls a web search tool, reads pages, keeps notes in a filesystem and returns a structured, cited report; long tasks run on create_deep_agent with subagents. It must never present a claim without a source that actually supports it.Knowledge work / Research
• Multimodal document understanding for KYCA LangChain v1 agent reads ID cards and address proofs (images and PDFs) with a vision-capable chat model, extracts fields into a strict schema, verifies them with deterministic tools, and pauses for a human reviewer on any mismatch. It must never auto-approve a case with a mismatch, and it must never invent a field it could not read.Fintech / Banking
• Pull request review assistantA LangChain v1 agent built with create_agent reviews each pull request against a repo checklist, using tools to fetch the diff and files, run linters and post review comments. It must never request changes or post a blocking review without a human approving it through HumanInTheLoopMiddleware.Software engineering
• Support FAQ bot with semantic cachingA telecom FAQ bot that checks an embeddings-based semantic cache before running a LangChain RAG agent, purges cached answers when knowledge-base articles change, and hands low-confidence or account-specific chats to a human. It must never serve a cached answer that belongs to a different question, is based on an outdated article, or contains another customer's personal data.Telecom
• CRM sales copilotA LangChain v1 agent that answers account questions from live CRM data through read tools, drafts follow-up emails, and proposes activity logs and field updates as structured output. It must never write to the CRM, or send anything, without explicit approval from the rep.B2B Sales / SaaS

LangGraph

Stateful agent graphs: human approval, supervisors and long-running workflows.

• Stateful graph with human reviewNodes, conditional edges and a checkpointer turn an agent into a graph that can loop, pause and resume.Banks & Regulated industries
• Multi-agent supervisorOne supervisor agent routes a task between specialist agents and decides when it is done.Research & Consulting
• Customer support agent that resolves refundsA LangGraph agent answers order questions, applies the refund policy and hands big refunds to a person.E-Commerce
• Prior-authorization assistantAn agent assembles insurer approval requests from the medical record, criterion by criterion.Healthcare
• Emergency triage & bed allocationAn agent suggests acuity, checks live capacity and drafts the handoff; nurses make every call.Hospitals
• Motor insurance claims processingDocuments, photos and policy rules are checked in parallel, and only risky claims reach an adjuster.Insurance
• On-call incident response agentAn investigator agent reads logs, deploys and runbooks, then proposes a fix for the engineer to approve.Tech industry
• Supply-chain disruption plannerWhen a route breaks, agents find the affected orders, price the options and wait for a planner's choice.Logistics & Retail
• Plan-and-execute agent with replanningA LangGraph StateGraph where a planner writes a step list, an executor runs one step at a time with tools, and a replanner rewrites the remaining plan from results until it returns a final answer. It must never lose or duplicate past step results, and it must never loop without a hard stop.Business analytics
• Self-correcting code generation loopA LangGraph StateGraph generates code, runs its tests in an isolated sandbox, reflects on failures and retries within a fixed attempt budget. A conditional edge routes each result to pass, retry or give up, and a Postgres checkpointer saves every attempt. It must never ship code that did not pass the real tests, and it must never loop without a bound.Developer tooling
• Travel booking with agent handoffsFlight, hotel and payments agents in one LangGraph StateGraph pass control to each other with Command(goto=..., update=...) over a shared typed state, and a human approves the final itinerary through interrupt() before any money moves. It must never charge the traveller twice, and never charge an amount or itinerary the traveller did not explicitly approve.Travel
• Personal assistant with long-term memoryA LangGraph assistant that recalls a user's saved facts and preferences from a namespaced store using semantic search, keeps each conversation's state with a checkpointer, and runs a memory-writing node after conversations to decide what to save. It must never leak one user's memories into another user's session, and it must never keep a fact the user asked it to forget.Consumer software
• Loan underwriting with durable executionA LangGraph StateGraph takes a loan application through document intake, bureau check, income verification and risk scoring, with a Postgres checkpointer so a crashed run resumes from its last checkpoint and an underwriter decides edge cases through interrupt(). It must never repeat a side effect on resume, such as a second bureau pull or a second booking, and must never issue a decision without a recorded reason and reviewer.Lending / Fintech
• Parallel summarisation with the Send APIA LangGraph map-reduce graph loads a batch of documents, such as quarterly reports, fans out one Send per document to summarise them in parallel, and merges the results with an operator.add reducer. A reduce node then writes one final brief. It must never silently drop a document or attribute a figure to the wrong report.Financial research
• Debugging an agent with time travelPulls a misbehaving support-agent thread from its checkpointer, inspects the state history, replays from an earlier checkpoint and forks it with update_state to test a different tool result, then compares outcomes to find and lock in a fix. It must never let a replay or fork trigger real side effects, such as refunds or customer emails, against production systems.Customer support
• Customer onboarding (KYC) with subgraphsA LangGraph parent graph runs identity, address and sanctions-screening subgraphs in parallel, each with its own private state, then merges their results and either approves the applicant or pauses for a compliance analyst. It must never open an account when any check failed, was inconclusive or produced an unresolved sanctions hit.Banking and fintech
• Streaming tutor appA LangGraph tutoring agent streams LLM tokens (messages mode), step progress (updates mode) and hint cards (custom mode) to a web UI, and persists each student's session as a checkpointed thread. It must never mix one student's thread or memory with another's, and must never leak internal node output, such as grading notes, into the chat stream.Education
• Corrective (adaptive) RAGA LangGraph StateGraph routes each question, retrieves and grades documents, rewrites the query or falls back to web search when retrieval is weak, then checks that the generated answer is grounded before returning it. It must never return an answer that is unsupported by the retrieved or searched context, and it must never loop without a bound.Enterprise knowledge
• Deep research agent with subagentsA long-running Deep Agents research agent on LangGraph plans with a todo list, delegates topics to subagents with isolated context, keeps notes in a filesystem and compiles a cited report. It must never state a claim in the final report that is not backed by a source the agent actually fetched.Knowledge work

LangFlow

Visual flows for chatbots, helpdesks and lead qualification.

• A visual RAG chatbot in LangflowDrag components onto the canvas, wire their ports, test in the Playground, then call it as an API.Corporate teams
• HR & IT helpdesk agentA Langflow flow answers policy questions, runs approved IT actions and files tickets from Slack or Teams.Corporate
• Property lead qualificationA Langflow flow reads each enquiry, checks the CRM and routes hot buyers to a site visit.Real estate
• Guest review triageEvery review is tagged, urgent ones reach the duty manager, and the rest get an approved reply.Hospitality
• PDF Q&A flow served as an APITwo Langflow flows: one ingests uploaded PDFs by splitting, embedding and storing chunks in a vector store, and one answers chat questions from retrieved context. A web app calls both through POST /api/v1/run/<flow-id> with tweaks. The system must never answer from another tenant's documents and must never present ungrounded text as if it came from the PDF.EdTech / Knowledge tools
• Agent component with toolsA Langflow Agent answers operations questions by choosing among a calculator, URL fetch, API Request and a runbook flow used as a tool, and it is tested in the Playground before being called over the API. It must never invent numbers or system status that did not come from a tool call, and it must never take a write action against production systems.IT operations
• Exposing flows as MCP toolsPublishes tested Langflow flows as tools on Langflow's MCP server so IDE assistants and desktop MCP clients can discover and call them with clear descriptions, API-key authentication and versioned names. It must never expose an unauthenticated endpoint or silently change a tool's behaviour under the same name.Developer tooling
• Multi-agent content pipelineA Langflow flow chains researcher, writer and editor agents to turn a topic into a blog draft. A separate webhook-triggered flow publishes it only after a human approves; nothing reaches the CMS without that sign-off.Media & Marketing
• Webhook-triggered ticket enrichmentA helpdesk webhook triggers a Langflow flow that classifies each new ticket, extracts key fields, looks up the customer, and writes a priority and a draft reply back to the ticket. It must never attach the wrong customer, overwrite a priority an agent has already set, or send a reply to the customer without a human approving it.Customer support
• Chatbot with per-session memoryA Langflow restaurant reservations chatbot that keeps chat memory per session id, follows a host-style instruction prompt, and calls tools to check availability and create bookings. It must never mix one guest's conversation into another's, and must never confirm a table that the reservations system has not actually booked.Hospitality
• Structured extraction from emails into a databaseOrder emails arrive through a Langflow webhook, are cleaned, turned into schema-shaped JSON by a model, validated in code, and upserted into an orders database, with every failure sent to a queue for review and replay. It must never insert an unvalidated or duplicate order, and it must never silently drop an email.E-Commerce / B2B distribution
• Building a custom componentA developer writes a Langflow custom component as a Python class that wraps a company API with typed inputs and outputs, tests it, and shares it so every flow and Agent uses the same version. It must never leak the API key, hide API failures as empty successes, or let flows drift onto different copies of the code.Enterprise software
• Document ingestion pipelineA scheduled Langflow flow that loads changed source files, attaches filterable metadata, splits them into chunks, embeds them and upserts them into a vector store, so a RAG knowledge base stays current. It must never leave stale or duplicate chunks, and it must never drop the metadata that retrieval filters depend on, such as access tags.Enterprise knowledge management
• Marketing copy with prompt chainingA Langflow flow turns a structured product brief into benefits, headline variants and channel-specific ad copy through chained Prompt and model components, then gates every line through a brand-rule check. It must never ship copy that breaks brand rules, makes unapproved claims or exceeds a channel's length limit.Marketing
• Taking a flow to production with tracingPackages a tested Langflow flow into a Docker deployment behind API-key authentication, with secrets loaded from environment variables, every run traced to LangSmith or Langfuse, and latency and errors monitored. It must never expose an unauthenticated run endpoint or leak provider keys into flow files, logs or traces.AI platform engineering

LangSmith

Tracing, evaluation and feedback loops that keep AI features improving.

• Trace, evaluate, improveLangSmith closes the loop: trace every run, turn real traffic into datasets and score each change.Tech & SaaS
• Release gate for a coding agentGolden tasks, CI and an LLM judge decide whether a new agent version is good enough to ship.Software & Tech
• Prompt & model experimentsVersioned prompts, one dataset and side-by-side experiments decide which prompt and model ship.EdTech
• Customer feedback to regression testsThumbs-down answers are routed to reviewers, corrected, and kept as tests for every later release.Banking
• Evaluating a RAG applicationRuns a versioned dataset of questions with reference answers through several RAG configurations using LangSmith evaluate(), then scores each run for correctness, groundedness and retrieval relevance with LLM-as-judge evaluators. It must never declare a winning chunk size or retriever from scores on different dataset versions, single noisy runs or a judge nobody has checked against human labels.Software / AI applications
• Calibrating LLM-as-judge evaluatorsBuilds an LLM-as-judge prompt with few-shot examples, measures its agreement with human labels collected in a LangSmith annotation queue, and iterates until it matches reviewers before it is trusted in production. It must never report judge quality using examples the judge has already seen as few-shots, and never ship a judge that has not been checked against humans on each important slice.AI / LLM engineering
• Pairwise comparison of two app versionsRuns the baseline and candidate app versions as two experiments on the same pinned dataset, scores them head to head with a pairwise LLM judge and a pairwise annotation queue, and gates the release on the result. It must never compare outputs from different dataset versions or let answer position bias the verdict.AI product engineering
• Online evaluation & production alertsScores sampled production runs and threads with online evaluators, shows quality, latency and cost on dashboards, and uses alerts and automation rules to route bad runs to human review and regression datasets. It must never let a quality or cost regression go unseen, and never blow the judge budget by scoring every trace.AI / LLM applications
• Human review with annotation queuesRoutes selected production runs into a LangSmith single-run annotation queue where several reviewers, each holding a reservation, score them against a rubric and assertions. Approved runs are then exported to a dataset for regression experiments. The workflow must never let unreviewed or disputed runs become reference outputs, and it must never let two reviewers silently overwrite each other's work.AI / LLM operations
• Prompt versioning in the Prompt HubEngineers edit prompts in the Playground, commit immutable versions to the Prompt Hub, promote them by moving staging and prod tags, and have applications pull prompts by tag so a rollback is a tag move, not a redeploy. It must never serve an unevaluated or unknown prompt version to production, and every production trace must record which commit produced it.AI / LLM platforms
• Evaluating agent trajectoriesBuilds LangSmith datasets of support tickets with reference tool-call sequences. Each agent version is scored on the path it took (agentevals trajectory match and trajectory LLM-as-judge) and on the final answer. It must never pass an agent that reaches a plausible reply through a wrong or unsafe tool path, such as issuing a refund before verifying the order.Customer support
• Cost & latency monitoring from tracesInstruments an LLM app so every LangSmith run records token usage, cost and latency, then uses trace views, dashboards by model and feature, and alerts to find slow steps and catch cost spikes. It must never misattribute cost or latency to the wrong model, feature or step, because teams act on these numbers.AI / SaaS
• Tracing any app with OpenTelemetryInstruments a non-LangChain, multi-service LLM app with @traceable, wrap_openai and the OpenTelemetry SDK. Spans go over OTLP to a collector that fans them out to LangSmith and an infrastructure backend, so one request yields one connected trace in both. It must never break trace context between services, and it must never ship secrets or raw PII to either backend.Developer tooling
• Curating golden datasets from productionTurns live LangSmith traces into a versioned, split golden dataset by sampling representative and failing runs, human-correcting reference outputs, adding synthetic edge cases and retiring stale examples. It must never let unreviewed outputs or PII become reference answers, and experiments must always pin an exact dataset version.AI / LLM applications
• Evals in CI with pytest and VitestLLM test files in pytest (@pytest.mark.langsmith) or Vitest (langsmith/vitest) run on every pull request. Each run syncs to LangSmith as an experiment on a dataset built from the tests, the build fails on regressions, and reviewers inspect results in the UI. The gate must never pass a real quality regression, and it must never block merges on judge noise or flaky model calls.AI software engineering

Companies

• Google: Search & Ads, and how Google shipsA search query flows through Google's front end, query understanding, sharded index fan-out, ranking and a parallel ads auction to return a results page. Click logs then feed quality evaluation and experiments. The system must never return slow or irrelevant results, and it must never let ads masquerade as or corrupt organic ranking.Internet search and advertising
• Amazon: order to doorstep, and working backwardsTraces an Amazon order from cart and payment through fulfilment-centre selection, pick-pack, sortation and last-mile tracking, alongside the publicly described operating practices of PR/FAQ, six-pagers, two-pizza teams and weekly business reviews. It must never double-charge, oversell stock or break the promised delivery date shown at checkout.E-Commerce
• Microsoft: Teams & Azure, and how Microsoft shipsAn end-to-end view of a Microsoft Teams meeting and chat on Azure, covering Entra ID sign-in, presence, chat, real-time media, recording and Copilot, alongside Microsoft's publicly described practices for planning, ring-based safe deployment, incident management and the Security Development Lifecycle. The workflow must never leak data across permission or tenant boundaries, and must never push a bad change everywhere at once.Enterprise software / Cloud
• Meta: feed & ads, and how Meta shipsWhen a person opens Facebook or Instagram, the system gathers candidate posts, ranks them with ML, inserts ads chosen by an auction, filters content through integrity checks and logs every interaction to retrain models. Small teams ship it continuously behind feature flags and A/B tests. It must never show violating content or a broken experience to everyone at once.Social media & advertising
• Anthropic: from model training to the Claude APITraces how a Claude model moves from data and pretraining through RLHF and Constitutional AI, Responsible Scaling Policy safety gates and staged release, to the Claude API serving requests with rate limits, prompt caching and usage monitoring. It must never ship a model whose measured capabilities exceed the safeguards in place, and it must never let one tenant's traffic or cached context affect another's.AI / Foundation models
• OpenAI: serving ChatGPT, and how OpenAI shipsTakes a ChatGPT or API message through auth, history, safety checks, model routing, streamed GPU inference, tools and usage metering. On the operating side it covers how models are trained, evaluated, safety-reviewed and rolled out in stages. It must never leak one user's data to another, skip safety checks, or ship a model change that has not been evaluated.AI / Large language models
• Apple: App Store to iPhone, and how Apple shipsTraces an iOS app from Xcode build and code signing through App Store Connect, TestFlight, App Review and CDN delivery to an iPhone, plus in-app purchases, updates and iCloud sync, alongside Apple's publicly described annual release and privacy-review practices. It must never deliver an unsigned or tampered binary, grant an unverified purchase, or expose user data beyond what the user consented to.Consumer technology / Mobile platforms
• Salesforce: lead to cash on a multi-tenant CRMA lead from a web form is scored, routed, converted to an opportunity, quoted, closed into an order and handed to service, all inside one shared multi-tenant platform customised through per-org metadata. It must never leak data across tenants or let one org's code or load degrade another's.Enterprise SaaS / CRM
• Flipkart: Big Billion Days, and how Flipkart shipsAn end-to-end flash-sale path for The Big Billion Days, from edge admission and a virtual queue through inventory reservation, UPI/EMI/COD checkout, fulfilment-centre allocation and Ekart delivery, plus the months of readiness work behind it. It must never oversell stock, charge a customer twice, or lose a paid order.E-Commerce

n8n

• Lead capture, enrichment & AI scoringA website form posts to an n8n Webhook, the lead is validated and enriched over HTTP, a model reached through Vercel AI Gateway scores and summarises it, and an IF node alerts Slack for hot leads while every lead is upserted into the CRM. It must never lose a lead, create duplicate CRM records or let a malformed AI response decide routing.B2B Sales & Marketing
• Support email triage with AI draftsAn n8n workflow reads new support emails from Gmail, classifies intent and urgency with a model reached through Vercel AI Gateway, routes each email with a Switch node, saves AI-written replies as Gmail drafts for a human to send, and pages urgent tickets to Slack. It must never send a reply to a customer without a person approving it, and it must never let an urgent email go unflagged.Customer support
• Company docs chatbot with the AI Agent nodeEmployees ask questions in an n8n chat, and an AI Agent answers from company documents using a vector store tool, per-session memory and a chat model reached through Vercel AI Gateway, while a separate scheduled workflow keeps the vector index fresh. It must never present an answer as company policy unless retrieved documents support it, and it must never mix one user's chat history into another user's session.Internal tools / Knowledge management
• Invoice extraction & approvalInvoices arriving by email are parsed by an n8n Information Extractor (model via Vercel AI Gateway), checked in a Code node, held in a Wait node for manager approval, then posted to accounting and Google Sheets. It must never post an invoice that is unapproved, duplicated or has totals that do not add up.Finance operations
• News to social posts pipelineA scheduled n8n workflow reads RSS feeds, drops items it has already seen, drafts channel-specific posts through Vercel AI Gateway, holds them for human approval in Slack, then publishes and logs each result. It must never publish a post that no one approved, and it must never post the same story twice.Media and marketing
• Meeting recordings to notes & tasksAn n8n workflow takes a recording-upload webhook, transcribes the audio, extracts a summary and action items through Vercel AI Gateway, creates tasks in the project tool and emails the summary to attendees. It must never create duplicate tasks, assign items to people who were not in the meeting, or email notes to anyone outside the invite list.Workplace productivity
• E-commerce order sync & stock alertsAn n8n workflow receives store order webhooks, decrements inventory in Postgres line by line using Loop Over Items, and raises WhatsApp/Slack alerts plus a purchase draft when stock crosses its reorder point. It must never double-count an order or silently drop a failed update.E-Commerce
• Employee onboarding automationAn n8n Form Trigger collects new-hire details, and an orchestrator workflow calls sub-workflows through Execute Workflow to create accounts, send documents for e-signature, book intro meetings and draft a personalised welcome message through Vercel AI Gateway. It must never create duplicate accounts, grant access before it is due, or send a welcome message that HR has not approved or that contains wrong facts.Human resources
• Incident alert routing with AI summariesn8n receives monitoring webhooks, drops duplicates, pulls recent logs over HTTP, asks a model via Vercel AI Gateway for a short summary with likely causes, and routes by severity to on-call paging and a Slack status channel. It must never drop or delay a critical page, even when log fetching or the model fails.Software operations
• Resume screening with human approvalAn n8n workflow takes job applications from an n8n Form, extracts resume text, scores it against the job through a model on Vercel AI Gateway, and pauses on a Wait-for-form step until a recruiter approves each shortlisted candidate. Only then does it send interview invites. It must never send an invite, or a rejection, on the model's judgement alone.Recruitment / HR
• Scheduled ETL between systemsA nightly n8n workflow pulls changed records from a paginated source API, transforms and validates them in Code nodes, upserts them into a Postgres warehouse in batches, and posts a run report, with an Error Trigger workflow for failures. It must never lose or duplicate rows: reruns must be idempotent, and the sync cursor may only move forward after a successful load.Data engineering
• Appointment booking agent on WhatsAppPatients message a clinic on WhatsApp. An n8n AI Agent, using a model through Vercel AI Gateway, checks slots, books, reschedules or cancels appointments, and hands off to staff when needed. It must never confirm an appointment that was not actually written to the calendar, and it must never double-book a slot.Healthcare