Domain 22: System Design Case Study Bank
ML and LLM system design: a design method, then case studies on feeds, search, ads, fraud, enterprise RAG, ChatGPT-scale serving, gateways and agent platforms.
Unlock all 55 lessons with Unrolled Pro
Full production code, quantitative capacity math, failure recovery playbooks, and interview coaching.
Lessons
55 lessonsThe System Design Method — From a Vague Prompt to a Defensible ML System at Staff Level
'Design a system that recommends the next lesson to each Atlas student.' Eight words, an hour, and a whiteboard.
Capacity Estimation and Sizing — Back-of-the-Envelope Numbers for ML Systems at Staff Level
Before building Atlas's AI tutor for a whole university, someone has to ask: how many students at once, how many tokens per answer, how many accelerators, how much a month?
Feature Stores and Training–Serving Parity — Point-in-Time Data, Freshness Tiers and Skew You Can Measure at Staff Level
Atlas's 'next lesson' model was trained on a nightly table where 'lessons completed this week' was counted one way, and served by an API that counted it another way.
Case Study: Fraud and Abuse Detection — Cost-Based Decisions, Real-Time Features and Delayed Labels at Staff Level
Atlas's paid tier attracts stolen-card payments, account-sharing rings selling certificates, and bots farming free credits.
Case Study: Spam Filtering — Tiered Actions, Cascades and Campaigns at Staff Level
Atlas's study forums and direct messages attract link spam, fake 'answer key' sellers and bot accounts.
Case Study: Churn Prediction — Clean Labels, Uplift and Measured Interventions at Staff Level
Atlas loses about 3% of its paying students every month.
Case Study: Surge Pricing — Forecasts, a Stable Price Controller and Marketplace Experiments at Staff Level
Atlas's live-tutoring marketplace has the same shape as ride-hailing: on exam evenings, students requesting a tutor outnumber tutors online, wait times climb, and both sides leave.
Case Study: ETA Prediction and Routing — Live Traffic, Graph-Plus-ML Correction and Honest Promises at Staff Level
When Atlas's live-tutor queue tells a student 'a tutor will join in about 4 minutes', it is making an ETA promise: a number people plan around, where being late costs more than being early.
Case Study: Sensor Anomaly Detection and Predictive Maintenance — Edge Features, Normal-Behaviour Models and Alert Budgets at Staff Level
Atlas runs its own GPU fleet for tutoring and grading. A disk that starts logging read errors, or a GPU whose temperature creeps up under the same load, often gives days of warning before it fails.
Case Study: Credit Underwriting — Delayed Labels, Explainable Models and Fair, Economic Decisions at Staff Level
Atlas offers students a pay-in-instalments option for certificate programmes.
Case Study: Patient Readmission Risk — Leak-Free Labels, Capacity-Aware Targeting and Proving the Care Helped at Staff Level
Atlas's student success team can call only so many learners a week.
Case Study: Energy Load Forecasting — As-Issued Weather, Per-Horizon Models and Honest Intervals at Staff Level
Atlas reserves GPU capacity a day ahead, and exam weeks bring sharp evening peaks. Reserving too little means slow tutors at the worst moment; too much is wasted money.
Case Study: Ad Click-Through Rate Prediction — Calibration, Freshness and the Auction at Staff Level
Atlas's course catalogue carries a few clearly labelled sponsored slots from partner universities.
Case Study: Recommendations and Search on One Platform — Shared Retrieval, Graded Labels and Bias Correction at Staff Level
Atlas needs a home feed of next lessons, search over 120,000 lessons and notes, and a few labelled sponsored placements from partner universities.
Case Study: Short-Video Recommendation Feed — Fresh Retrieval, Multi-Task Ranking and Long-Term Guardrails at Staff Level
Atlas's home screen shows a feed of short lesson clips. A feed that maximises minutes watched today can drift toward entertaining but useless clips and lose learners next month.
Case Study: E-commerce Search and Substitution — Availability-Aware Retrieval, Learned Ranking and Constraint-Safe Replacements at Staff Level
Atlas's course search must understand 'pytorch for beginners, weekend schedule' and, when a live cohort is full, suggest an equivalent one without breaking the learner's constraints: same language, same level, a time they can attend.
Case Study: Music Discovery Playlists — Taste Clusters, Cold-Start Audio and Whole-Playlist Design at Staff Level
Every Monday, Atlas could send each learner a short 'what to learn next' list: lessons they have not seen, close enough to what they already know to feel reachable, spread across their interests, with one or two stretches.
Case Study: Job and Candidate Matching — Mutual Fit, Congestion Control and Fair Hiring Signals at Staff Level
Atlas connects graduates of its programmes with employers looking for those skills.
Case Study: Friend Suggestions on a Social Graph — Pruned Candidates, Meaningful-Connection Labels and Privacy by Design at Staff Level
Atlas suggests study partners: classmates in the same course and time zone, people who joined the same study group.
Case Study: News Feed Ranking — Hybrid Fan-Out, Value Models and a Separate Integrity Layer at Staff Level
Atlas's community home page mixes posts from classmates, study groups and instructors.
Case Study: Vacation Rental Search — Availability as a Hard Filter, Booking-Aware Ranking and a Fair Marketplace at Staff Level
Atlas lists in-person study retreats hosted by partner campuses: a few rooms on specific dates, searched by place and date, with hosts who can accept or decline.
Case Study: Food Delivery Dispatch — Predicting Preparation, Batch Matching and Fair Offers to Couriers at Staff Level
When a student presses 'get live help', Atlas must pick a mentor: not just whoever is free right now, but someone free when the student has finished describing the problem, with the right skills, without leaving other mentors idle or overloaded.
Case Study: Search Autocomplete — Precomputed Prefixes, Fresh and Personal Ranking, and Suggestions That Are Safe to Say at Staff Level
When a learner types 'gradient d' into Atlas's search box, suggestions should appear before the next keystroke: 'gradient descent', 'gradient descent with momentum'.
Case Study: Visual Search — Routing, Embeddings Across the Photo Gap, Compressed Indexes and Privacy Rules at Staff Level
Atlas lets learners photograph a handwritten equation, a textbook diagram or a whiteboard sketch and find the lesson that explains it.
Case Study: Real-Time Ad Bidding — Valuation in Milliseconds, First-Price Shading and Budgets That Never Overspend at Staff Level
Atlas buys ads on ad exchanges to reach people who might want to learn AI.
Case Study: Question Deduplication — Intent-Level Matching, Tiered Actions and Humans in Charge at Staff Level
Atlas's forums see the same questions every cohort: why the loss is NaN, why the gradient vanishes, how to choose a learning rate.
Case Study: Dating App Matchmaking — Reciprocal Ranking, Attention Balancing and Safety First at Staff Level
Atlas pairs learners for weekly mock interviews: both people must want the pairing, a few popular mentors cannot take every request, and people share sensitive things about their careers.
Case Study: App Store Search — Navigational Intent, Retention-Based Labels and Integrity Against Gaming at Staff Level
Atlas has a growing library of community-built practice notebooks and plugins.
Case Study: An LLM Assistant Platform — Token Planning, Routing, Grounding and Safe Actions at Staff Level
Three partner universities want Atlas inside their own walls: answers grounded in their private course material, strict access control per course, their own cost budgets, and agents that can file extensions and book office hours.
Case Study: Multi-Tenant Enterprise RAG — Permissions, Freshness and Cost at Staff Level
Universities want Atlas answering from their own private course material, with every student seeing only what their enrolments allow.
Case Study: Inline Code Completion Engine — Latency, Context and GPU Cost at Staff Level
Atlas's coding exercises want ghost-text hints as students type.
Case Study: Customer Support Agent — Actions, Policy and Hand-off at Staff Level
Atlas students email about extensions, refunds and lost certificates. An agent that can look up their account and act within policy would resolve most of these in minutes.
Case Study: Legal Document Analysis — Extraction with Evidence, Recall and Confidentiality at Staff Level
Universities sign thousands of licensing and partnership agreements, and someone has to know which ones auto-renew, which forbid data leaving the country, and which end if the university merges.
Case Study: Clinical Q&A with Citations — Grounding, Safety Tiers and Patient Privacy at Staff Level
Atlas's health-sciences students practise clinical reasoning against real guidelines.
Case Study: Financial Filings Summarizer — Exact Numbers, Change Detection and Minutes-After-Release at Staff Level
Atlas's finance track teaches students to read company reports. Professionals want the same skill as a service: the moment a company publishes its results, a summary of what changed, with every number exactly right.
Case Study: Multimodal Voice Assistant — Turn-Taking, Device–Cloud Split and Safe Actions at Staff Level
Atlas learners practising interview answers want to speak, not type, and get a natural spoken reply.
Case Study: Autonomous Software Engineering Agent — Sandboxes, Verification and Cost per Task at Staff Level
Atlas's own codebase has a backlog of small issues: broken links, flaky tests, outdated dependencies.
Case Study: Text-to-SQL Analyst — Semantic Layers, Correctness and Warehouse Cost at Staff Level
Atlas's team asks questions like 'which lessons have the highest drop-off for learners who joined last month?' and waits days for an analyst.
Case Study: Long-form Document Synthesis — Planning, Consistency and Claim-level Citations at Staff Level
Atlas's curriculum team writes 30-page reviews of how a topic is taught across dozens of sources.
Case Study: Multi-agent Research Engine — Orchestration, Token Budgets and Web Safety at Staff Level
Learners ask Atlas open questions like 'how do production teams evaluate RAG systems today?' A good answer means searching many sources in parallel, reading them critically and writing a cited synthesis.
Case Study: Web Automation Agent — Perception, Irreversible Actions and Injection Defence at Staff Level
Atlas learners juggle conference registrations, visa forms and course enrolments across clunky websites.
Case Study: Email Assistant with Drafts — Triage, Voice and Inbound Injection at Staff Level
Atlas instructors answer hundreds of learner emails a week: extension requests, grading questions, scheduling.
Case Study: Meeting Copilot with Actions — Transcription, Attribution and Consent at Staff Level
Atlas study groups meet online every week and forget half of what they agreed.
Case Study: High-throughput LLM Gateway — Token Quotas, Failover and Overhead at Staff Level
Every Atlas feature (tutor, grader, search, practice generator) calls language models from several providers.
Case Study: Multi-model Semantic Router — Cost–Quality Routing, Shadow Evaluation and Drift at Staff Level
Most Atlas tutor questions ('what does this error mean?') are easy, and a few ('why does my proof of convergence fail here?') are hard.
Case Study: Multi-tenant LoRA Serving — Adapter Batching, Memory Tiers and Isolation at Staff Level
Each partner university wants an Atlas tutor tuned to its own courses and tone. Training a full model per university is far too expensive, but small adapters (LoRA) on one shared base model are cheap to train.
Case Study: Prefix Caching KV Store — Block Keys, Prefix-Aware Routing and Memory Tiers at Staff Level
Every Atlas tutor request starts with the same long system prompt, then the course material, then the student's conversation so far. Only the last message is new.
Case Study: Speculative Decoding Engine — Acceptance Rates, Lossless Verification and Load-Aware Control at Staff Level
Atlas tutors answer in long, step-by-step explanations, and students watch the words appear. Generation speed per student is what they feel.
Case Study: LLM Guardrail and Safety Gateway — Check Cascades, Streaming Moderation and Failure Policy at Staff Level
Atlas serves students of all ages, teachers and university staff, and its tutors read course documents and call tools.
Case Study: Real-time Vector Database at Scale — Segments, Filtered Search and Deletion Guarantees at Staff Level
Atlas retrieval searches across course materials, student notes and forum posts from many universities.
Case Study: LLMOps Evaluation Gates — Release Bundles, Statistical Gates and Canaries at Staff Level
Atlas ships changes every day: a new tutor prompt, a different retrieval setting, an upgraded model, a new tool.
Case Study: Synthetic Data and Distillation — Prompt Coverage, Quality Filters and Escalation at Staff Level
Atlas grades millions of short practice answers a day and drafts feedback for each one. A large model does it well but costs ten times more than a small one.
Case Study: Privacy-preserving LLM Gateway — Data Classes, Routing, Pseudonymisation and Retention at Staff Level
Atlas staff use language models for everything from drafting course announcements to summarising student support cases, and some of those cases contain grades, health accommodations or contact details.
Case Study: Geo-distributed LLM Serving — Residency Islands, Shared GPU Pools and Regional Failover at Staff Level
Atlas has students in the Americas, Europe and Asia, and some universities require that their students' data never leaves their region.
Capstone: Atlas at 10 Million Students — One Design to Fund, Run and Grow at Staff Level
The final brief: Atlas goes from 400,000 to 10 million students in two years, across four regions and three partner universities, with tutoring, voice, recommendations, grading and agents.