Systems, not screenshots.

System brief

The backend supports customers, cashiers, and administrators without collapsing their permissions into one controller. Loyalty commands pass through authentication, rate policy, idempotency checks, and service rules before state changes are accepted.

Key engineering decisions

• Safe retriesIdempotency handling protects transaction-style commands from repeated client requests.
• Abuse boundaryBucket4j token buckets apply per-IP and per-user policies, with explicit cleanup for inactive local buckets.
• Bounded workAn expired-redemption job processes a fixed batch every five minutes instead of attempting an unbounded sweep.
• VisibilitySpring Actuator, Micrometer, and Prometheus expose health and runtime metrics; request logging and audits provide operational context.

System brief

The application keeps its PostgreSQL connection server-only, scopes runtime behavior with a store identifier, and uses Neon as the shared source of truth. The same application build can run behind Vercel’s serverless runtime or as a Docker service on Render.

Key engineering decisions

• Secret boundaryThe database URL is never exposed through browser-prefixed environment variables.
• Portable runtimeA production Dockerfile and Compose definition match the cloud service contract instead of creating a separate local architecture.
• Preview isolationDeployment notes explicitly call for separate Neon branches and store identifiers when preview environments must not share production data.
• Platform healthThe Render blueprint defines a service health check and injects secrets at deploy time.

System brief

SafeBite scans products, queries specialized food, cosmetics, and medication sources, then maps the strongest available result into one shared model. A verdict service normalizes ingredient text, expands allergy aliases, and returns Safe, Warning, Danger, or Unknown with concrete reasons.

Key engineering decisions

• Parallel lookupThe scan orchestrator queries classification and detail sources concurrently, applies an eight-second bound, and selects medication, cosmetics, or food results by explicit priority.
• Graceful fallbackComposite clients and category-aware fallback paths keep one unavailable provider from collapsing the complete scan flow.
• Explainable verdictNormalized ingredient data is compared against saved allergy aliases and cosmetics watchlists; the result includes matched reasons rather than a black-box score.
• Local protectionAuthentication uses Firebase, profiles and history use Firestore, and sensitive preferences are stored through Android security crypto.

System brief

EduLife coordinates an Android app, a React web client, and a Spring Boot modular monolith. The backend owns permissions and the complete learner loop: discover, enroll, learn, take a server-scored exam, and receive a verifiable PDF certificate.

Key engineering decisions

• Pragmatic monolithDomain modules preserve clear ownership without introducing distributed-system overhead before the MVP requires it.
• Server authorityFirebase identifies the client, while the backend validates tokens, resolves internal users, enforces roles, and protects exam integrity.
• Database evolutionPostgreSQL and Flyway keep the learning model and schema changes explicit across courses, progress, attempts, analytics, and certificates.
• Container safetyThe multi-stage Java 21 image runs as a non-root user and applies container-aware memory limits for the deployed service.

System brief

Telemetry enters through an ingestion boundary, crosses Kafka, and fans out to processing and anomaly-detection consumers. PostgreSQL holds durable results, Redis supports fast access, and a dead-letter path is part of the resilience model.

Key engineering decisions

• DecouplingKafka separates ingestion pressure from independent telemetry and anomaly consumers.
• Local parityCompose provisions Kafka, Zookeeper, PostgreSQL, Redis, Prometheus, Grafana, and dedicated exporters with health checks and persistent volumes.
• Design recordArchitecture decisions document technology choices and the sequence from local Compose to later Kubernetes deployment.
• Honest scopeThe platform is presented as an active systems lab, not as a finished production service.