Use exact semantics and compatibility rules from sources such as RFC 9110, then state which revision and subset the design implements.
One discipline · five design lenses
Choose the system boundary. Then choose the method.
System design can mean a global backend, a processor and memory hierarchy, a real-time device, an entire multidisciplinary mission, or an analog AI accelerator. Start with what you are responsible for designing.
- 5design tracks
- 50+visual models
- 8interview prompts
# The word “system” is contextual
boundary = identify_responsibility()
if boundary == "services + data":
track = backend_distributed
elif boundary == "whole engineered product":
track = systems_engineering
elif boundary == "compute platform":
track = hardware
elif boundary == "physical feedback loop":
track = embedded
elif boundary == "analog AI accelerator":
track = npu_acim
state_requirements_and_tradeoffs()01 · Track selector
What are you designing?
Each track shares systems thinking, but its constraints, artifacts, bottlenecks, and interview language are different.
System boundaryWhich elements are inside your responsibility?
01 · Software services
Backend & Distributed
APIs, data stores, caches, queues, consistency, availability, traffic and operations.
Example: design a global news feed → 02 · Whole productSystems Engineering
Stakeholders, ConOps, requirements, allocation, coupled budgets, architecture search, trade studies, risk, integration and V&V.
Example: design an autonomous drone → 03 · Compute platformHardware & Computer
Workloads, microarchitecture, memory, interconnect, PPA, thermal and reliability.
Example: design an edge AI computer → 04 · Physical controlEmbedded & Cyber-Physical
Sensors, firmware, real-time deadlines, control, power, safety and actuators.
Example: design a smart camera → 05 · Analog AI computeNPU Software Stack · ACiM
Graph compiler, analog-aware mapping, runtime, calibration, drift and digital fallback.
Example: design a hybrid ACiM NPU →03 · Discipline decision graph
Follow the dominant design object
The same product may use multiple tracks. Choose the page matching the interviewer’s or team’s system boundary.
StartWhat must your architecture primarily coordinate?
Backend
Optimize throughput, latency, consistency and availability.
Open track →Systems Engineering
Manage requirements, interfaces, coupled budgets, architecture search, trade decisions, lifecycle risk and integration.
Open track →Hardware
Balance performance, power, area, cost and programmability.
Open track →Embedded
Meet deadlines, safety, power and control stability.
Open track →ACiM NPU
Preserve accuracy while exploiting efficient in-memory matrix operations.
Open track →04 · Integrated interview case
Design an offline-first autonomous inspection drone
A single product crosses every boundary in the Atlas. Practice turning mission intent into traceable contracts for flight control, edge compute, analog inference, fleet services, and end-to-end evidence.
Inspect safely when the network disappears
A battery-powered drone inspects thermal and visual anomalies inside a GNSS-denied industrial site. It must avoid obstacles and finish a 30-minute route without depending on the cloud, tolerate a 10-minute uplink outage, and synchronize authenticated observations when connectivity returns. The safety-critical flight loop stays on deterministic digital control; the ACiM NPU accelerates perception and must expose a monitored digital fallback.
- Mission outcomeCover the assigned assets, preserve evidence provenance, land safely, and report incomplete coverage rather than silently claiming success.
- Critical boundaryCloud services may plan, observe, and update missions, but loss of connectivity cannot remove local stabilization, obstacle avoidance, or safe-state authority.
- Evidence ruleEvery target is provisional until workload traces, timing analysis, power and thermal measurements, fault injection, and representative mission tests support it.
Artifact handoff graph
Make each discipline produce a contract the next one can consume
Read left to right for decomposition; follow test evidence back to the first requirement for closure.
- Systems EngineeringConOps, hazards, SHALL requirements, interface ownership, allocated budgets, modeled alternatives, feasible Pareto finalists, and the decision record.
- EmbeddedSensor-to-actuator timing chain, task and interrupt schedule, mode machine, health signals, and independently reachable safe state.
- HardwareWorkload envelope, memory and I/O paths, compute headroom, power rails, thermal limits, and observability hooks.
- NPU + ACiMSupported-op manifest, tile mapping, quantization and accuracy envelope, calibration identity, confidence telemetry, and fallback trigger.
- BackendVersioned mission and observation schemas, idempotent synchronization, rollout policy, device state, authorization, and fleet SLOs.
- Integrated evidenceTrace matrix, interface tests, latency and energy traces, drift tests, network-loss injection, safe-landing demonstrations, and residual risk.
Interface checkpoint: define units, ranges, timing, freshness, version, ownership, failure semantics, security, and observability for every arrow. “Send detections” is not a closed interface; a versioned detection schema with timestamps, coordinate frame, calibration ID, confidence meaning, retry behavior, and stale-data handling is.
| Design tension | Reasonable starting choice | Invariant / interface contract | Pressure test and evidence |
|---|---|---|---|
| Local autonomy vs cloud offload | Keep stabilization, avoidance, mission continuation, and landing local; use cloud services for planning, fleet learning, and delayed evidence processing. | Network loss cannot violate a safety deadline; reconnection cannot duplicate commands or observations. | Inject loss, delay, reordering, and duplicate delivery; measure local behavior and reconciliation correctness. |
| Perception accuracy vs ACiM efficiency | Map qualified operators to ACiM; retain a reduced digital path for unsupported operations, drift, or low-confidence output. | Calibration and model identities travel with results; fallback transition has a bounded latency and an explicit degraded mode. | Sweep temperature, drift, device variation, and hard scenes; compare task metrics and fallback coverage against the digital reference. |
| Response latency vs battery reserve | Use mode-aware rates: high-rate control, event-driven perception where justified, and batched noncritical synchronization. | Deadline and energy allocations have named owners and separate margin; thermal throttling cannot be ignored. | Replay worst-case sensor bursts while measuring deadline misses, rail power, temperature, and remaining mission reserve. |
| Evidence fidelity vs storage and uplink | Store authoritative event metadata and selected source frames locally; upload resumable chunks and derive previews separately. | Observation IDs are stable, hashes detect corruption, provenance survives retry, and retention follows mission policy. | Fill storage, interrupt uploads, corrupt a chunk, rotate credentials, and prove exact recovery without false completion. |
| Rapid updates vs assurance | Use signed, staged releases with compatibility gates and rollback for model, calibration, firmware, and backend schema changes. | An incompatible or unverified bundle cannot become flight-active; the fleet reports the complete active version set. | Canary representative hardware, force a rollback, and trace every changed artifact to the tests and requirements it invalidates. |
50-minute practice sequence
Move from mission to evidence without disappearing into one subsystem
Use the time boxes as interview checkpoints. Say which assumption you are making, which artifact records it, and what evidence could overturn it.
00–05 minFrame the mission Clarify users, environment, route, payload, offline duration, safety boundary, and what “inspection complete” means.
Checkpoint
State the system boundary, three success measures, three exclusions, and the highest-consequence failure.05–12 minWrite measurable requirements Translate the ConOps into functional, timing, energy, accuracy, availability, security, and evidence requirements.
Checkpoint
Assign provisional budgets and flag values that need measurement instead of presenting guesses as facts.12–20 minSketch the architecture Separate safety control, perception, storage, synchronization, fleet services, and human operations; draw trust and network boundaries.
Checkpoint
Walk one normal mission and one disconnected mission end to end, naming state owners and durable records.20–30 minClose critical interfaces Trace sensor-to-actuator latency, ACiM fallback, power and thermal limits, schema versions, calibration identity, and retry semantics.
Checkpoint
For each cross-track arrow, name producer, consumer, unit, deadline, version, fault response, and observable signal.30–39 minExpose bottlenecks and trades Test compute, memory bandwidth, bus contention, thermal throttling, battery reserve, storage pressure, and uplink backlog.
Checkpoint
Choose one alternative, quantify why it helps, and state the cost or new failure mode it introduces.39–46 minProve readiness & acceptance Build component, interface, hardware-in-loop, fault-injection, drift, security, operator and representative mission evidence upward.
Checkpoint
Connect each critical requirement to a method, configuration, acceptance rule and result; state who accepts the baseline, which risks remain open, and what must be ready before transition.46–50 minTransition, operate & retire Package the accepted baseline, hand off ownership and support, stage the signed rollout, define rollback or safe fallback, observe mission and fleet outcomes, and name the eventual decommissioning contract.
Checkpoint
Close with readiness owners, runbooks, rollout gates, rollback triggers, operating SLIs and alerts, the first signal that feeds the next improvement decision, and who safely ends authority, credentials, protected data and physical service.
05 · Interview practice matrix
Practice the right conversation for each role
Each track ends with a timed 45–55 minute practice arena, revealable coaching notes, pressure tests, and a self-score.
BackendURL shortener · feed · chatScale, storage, consistency, bottlenecks
SystemsAutonomous delivery droneConOps, allocation, optimization, interfaces, V&V
HardwareEdge AI computerWorkload, memory, PPA, thermal
EmbeddedBattery smart cameraReal time, control, power, safety
ACiM NPUHybrid analog AI stackCompiler, mapping, runtime, calibration