Tamir Yirga

Ski lift-ride events flow through RabbitMQ into DynamoDB, with Redis-cached reads. A fleet-aware admission controller divides rate limits across replicas and sheds load when the queue falls behind. Write batching cuts DynamoDB operations 96% (13,168 → 533) and HyperLogLog counting cuts distinct-value writes 62%.

Distinct-count write reduction (HyperLogLog)

62%  25,163 → 9,635

Fleet admission rate under saturation

99.8%

Defects documented / Architecture Decision Records

38 bugs · 11 ADRs

The 4,309 req/s figure is from an EC2 + production DynamoDB run. All post-rebuild comparisons ran on LocalStack locally. A batching throughput result that did not survive swapping run order is not on this page.

Audio goes through AWS Transcribe, then LLM extraction, into a 4-table PostgreSQL schema with JSONB for variable event shapes. A canonical event spec locks down field names and required fields per type across three teams. 14 PRs reviewed at the integration gate, 11 blocked on critical findings (cascade-deleting patient data, corrupted JSONB).

Flyway migrations, seven of them mine

8

Containers in the Compose stack

6

Mental state enum values the spec allows

12

Was data lead and integration reviewer in a 10-person team. Schema, migrations, event spec, and Compose stack. The extraction consumer, architecture diagrams, and model-selection ADR are teammates' work. No performance numbers were measured.

Compares ResNet-50, ConvNeXt-Tiny, EfficientNet-B3, and MobileNetV3 on the same GPU. MPS and CUDA queue work asynchronously, so the harness syncs after each of 100 timed passes to measure real GPU time. Reports mean, standard deviation, min, max, and derived throughput.

Architectures through same harness and hardware

4

Warmup passes before the clock starts

10

Timed passes per measurement

100

Three-person course project. Results are gitignored, no committed numbers to quote.

Each frame is thresholded via ISODATA over a 6.25% pixel sample, segmented into regions, and oriented by least central moment. Regions are described by 7 invariant features or projected onto 20 principal components from a 64×64 patch. One image teaches a new object because matching happens in that compressed 20-D space.

Principal components, down from 4,096 pixels

20

Rotation- and scale-invariant features per region

7

Pixels sampled to estimate the threshold

6.25%

No timing code committed, so no frame rate to quote. The demo video is the evidence.