Designs by Duhart · measured, not estimated
Every figure below came from a file census or a written report — never
an estimate. Vendored trees, virtualenvs, node_modules and generated
protobuf output are excluded, because counting them is the easiest way to make a
portfolio dishonest.
Single hue, more-is-darker in light and more-is-lighter in dark — magnitude only. Hover or focus a row for what it talks to.
| Service | Language | Lines |
|---|---|---|
| Gateway trio | Go | 48,582 |
| MatchService | Go | 33,638 |
| OpenHouseGlobal | TypeScript | 29,453 |
| GlobalChatService | Go | 26,176 |
| billboard-users | TypeScript | 10,581 |
| GlobalCloudflareAdapter | Go | 7,855 |
| Solomon retrieval | Go | 4,757 |
| BillboardActivityService | Go | 977 |
BillboardActivityService is 977 lines — the shortest service in this chart — and it sits at the top of the portfolio ranking. Line count measures effort, not leverage. A chart of size is a chart of size; it is not a chart of what will get you hired, which is why it appears here and not on the ranking page.
611,426 authored lines across 1,200 files, excluding Pods, frameworks and dependencies.
| Mode | Files | Lines |
|---|---|---|
| Dating | 147 | 90,161 |
| Array | 122 | 90,017 |
| Billboard | 128 | 57,669 |
| Main | 160 | 37,714 |
| BoxOffice | 45 | 20,407 |
| Universal | 26 | 17,530 |
| Global | 110 | 17,246 |
| OpenHouse | 57 | 9,294 |
| Identity | 51 | 8,051 |
| Registration | 13 | 6,072 |
| Vitals | 14 | 2,582 |
| Login | 5 | 1,427 |
| Storefront | 5 | 820 |
Tests, gates and measured results — the figures that are worth more than line counts.
56 Go (*_test.go), 39 Python, 58 TypeScript — spanning race tests, benchmarks, contract-coverage tests and live end-to-end runs.
Solomon's promotion pipeline. Two of the nine are latency and throughput, so "good enough" is defined on both axes before anything reaches a cluster.
Solomon Python, 62/62 passing — including a causal-mask leak test and an overfit-one-batch test.
Load test against the inference service: p95 2.7 s against a 6 s SLO, HPA scaling to 3 replicas.
The size of the Redis consumer-group PEL leak at the moment it was found in feed_index:writes.
Measured feature parity against Spotify, written before any remediation began.
F-01…F-49 in the Dating audit, each with file:line, severity and evidence.
Typecheck, unit, real end-to-end, contract diff, shadow types, secret detection — one of them added because of a specific past defect.
Chat history, match cards partitioned by (service_region, geohash4), feed timeline, OpenHouse. Schema migrates on boot.
The activity log (6 partitions, keyed by actor), gateway event transport, Solomon retrieval ingestion.
Presence, sessions, and a feed index using consumer groups — where the PEL leak lived.
The transactional outbox, notifications, profiles. Reached from the edge through Hyperdrive.
Edge-local relational storage for library and user data.
Audio and video objects, multipart ingest, HTTP range streaming.
IngestDO for single-writer upload sessions; JamzFingerprintDO running an FFT compiled to WASM.
Chat media on S3; configuration and hot lookups on KV — never anything that must stay ordered.