LesegoRabotapi

A serverless AWS uptime-monitoring platform built around scheduled checks, incident recording and event-driven alerting.

VigilWatch replaces an always-running monitoring server with an event-driven AWS architecture. EventBridge schedules health checks, Lambda executes them against external endpoints, DynamoDB records the resulting state, SNS handles alerts, and a static S3/CloudFront dashboard exposes the system through API Gateway and Lambda APIs.

Sole engineer · 2026

A Python ETL pipeline that turns raw public-transport route and stop data into a validated, deduplicated MySQL dataset.

MyInsights extracts raw public-transport route and stop records, standardises and deduplicates them with pandas, and loads the result into a relational MySQL schema. A route-ID dedup check that returned zero duplicates (while the database plainly held repeats) pushed the pipeline toward business-key deduplication instead, landing on 47 unique routes and 954 unique stops with SQL validation to back the counts.

Team of two

A cost-conscious AWS data lake that turns UNDP Human Development Index data into governed, queryable, partitioned analytics data.

The project takes raw HDI indicator data into an S3-based lake, uses AWS Glue to catalogue and transform it, and exposes the processed dataset through Athena for serverless SQL analysis. The design separates raw, processed and query-result data, layers Lake Formation over IAM for resource-level governance, and keeps the entire pipeline within AWS Free Tier constraints.

Sole engineer · archived

A local-first MLOps pipeline that validates, trains, serves and monitors a model, and correctly refused to ship a classifier its data couldn't support.

The project builds every stage of the ML lifecycle as testable, observable infrastructure: rule-based validation, scikit-learn feature pipelines, MLflow-tracked training, a containerised FastAPI inference service, and a Prometheus/Grafana stack, gated by GitHub Actions CI with Trivy image scanning. Hardened, imbalance-aware metrics then caught that the source dataset's label was statistically independent of its features, and the deliverable was reframed around that finding rather than a working classifier.

Sole engineer