/srv — Complete Work Inventory, Ranked for Hire
Owner: J. Michael Duhart · j.michaelduhart@gmail.com Compiled: 2026-08-27 · 28 projects Target role: Distributed Systems Engineer — Google / Spotify / financial institution Target comp: $210K, fully remote
How this is ranked
Every project in /srv is scored against what a distributed-systems hiring loop
at these three employers actually tests. Not "is it impressive" — is it the
thing they ask about on the whiteboard, and can you say "I built that and it
runs"?
| Weight | Signal | Why it matters to the target employers |
|---|---|---|
| ×5 | Canonical DS primitives — consensus, partitioning, replication, event logs, idempotency, fan-out, backpressure | This is the Google/Spotify systems-design interview. The whole loop. |
| ×4 | Production evidence — deployed, measured, has a runbook, has failure modes you can name | Separates "wrote a demo" from "carries a pager". Banks weight this highest. |
| ×4 | Correctness under concurrency — race conditions found, fixed, and tested for | Senior signal. The single hardest thing to fake in an interview. |
| ×3 | Scale artifacts — sharding keys, partition design, consumer groups, TTLs, hot-key handling | Spotify's backend and Google's SRE-adjacent roles probe this directly. |
| ×3 | Language fit — Go, Java, Python, C++ | Go is the lingua franca at Spotify's backend and Google Cloud. |
| ×2 | Contract discipline — OpenAPI/AsyncAPI/protobuf, CI-enforced | Banks require it. Google requires it culturally. |
| ×2 | Written communication — design docs that state trade-offs and costs | Google promo packets are made of this. |
| ×1 | Frontend / mobile / marketing | Real work, but it is not what this role hires for. |
Each project is rated 0–5 per signal. Max score 120. Ratings are a hiring judgment, not a measurement — treat differences under ~5 points as ties.
The ranking — all 28
Tier S — lead with these. They are the interview. (95+)
| # | Score | Project | One line | Where |
|---|---|---|---|---|
| 1 | 103 | Billboard Activity Graph | Transactional outbox → Kafka → three independent consumer groups → Cassandra + Postgres + realtime. At-least-once, idempotency paid explicitly at each sink. | app/BillboardActivityService |
| 2 | 106 | GlobalChat Realtime Messaging | 26K-line Go service: Cassandra chat, WebSocket fan-out, presence, push, calling signaling. Benchmarked and race-tested. | app/GlobalChatService |
| 3 | 101 | Match Service — Distributed Ranking | Geo-partitioned Cassandra ranking, Redis Streams consumer groups; a phantom-partition bug and a 952-message PEL leak, both fixed and written up. | app/MatchService |
The top three sit within 5 points of 120 — that is noise. GlobalChat edges the rubric on production evidence and breadth. The Activity Graph is listed first because it leads on the axis that decides a systems-design round: it is the canonical question, with every trade-off already written down. Open with the Activity Graph; bring GlobalChat as proof you have carried something this size in production.
Tier A — strong support. Infrastructure, ML platform, media. (75–94)
| # | Score | Project | One line | Where |
|---|---|---|---|---|
| 4 | 90 | Solomon — Serving, Kubernetes & Autoscaling | FastAPI SSE/WS streaming + continuous batching, k3s + Helm + HPA scaled to 3 replicas, NetworkPolicy, dry-run-validated manifests. | Solomon/k8s |
| 5 | 89 | Solomon — LLM Platform (umbrella) | The whole pipeline end to end. Decomposed into the six projects below. | Solomon/ |
| 6 | 84 | Billboard Production Hardening | 60+ commits from forgeable auth to RS256 four-tier, 9 D1 tables, resumable upload, moderation, DMCA, CI contract gates. | app/workers/billboard_users |
| 7 | 83 | WebRTC Calling Plane | Signaling/media-plane audit: ICE/STUN/TURN, P2P vs orphaned SFU, a fatal concurrent-map race located. | docs/webrtc-chat/ |
| 8 | 80 | Media & Upload Infrastructure | R2 multipart ingest across five modes, IngestDO single-writer sessions, TUS resumability, a Go adapter bridging edge to Kafka/Redis. | app/GlobalCloudflareAdapter |
| 9 | 79 | Push & Notification Pipeline | Kafka → consumer → APNs, plus PushKit VoIP; dedupe where a duplicate is a second buzz on someone's phone. | app/GlobalChatService |
| 10 | 75 | Solomon — Retrieval & Agent Tier | 4,757 lines of Go: chunk/embed/search/store, a bounded agent loop, a grounding guard with its own test, Kafka ingestion. | Solomon/retrieval |
Tier B — real systems work, narrower interview surface. (60–74)
| # | Score | Project | One line | Where |
|---|---|---|---|---|
| 11 | 73 | Platform Operations & CI/CD | systemd units including a restore-verify timer, Docker Compose stacks, two CI systems, gates that encode specific past defects. | deploy/, .github/ |
| 12 | 70 | Auth & Identity Platform | gRPC/Connect gateway, RS256 issuance, ACL, mode-sync, 24 protobuf contracts generating Go + Swift + TS. | app/AuthGateway |
| 13 | 70 | GlobalChat Messaging UI | iOS + web realtime clients with a transport/store/runtime split — and a wiring record that retracts three of its own conclusions. | IOS/.../GlobalChat |
| 14 | 69 | Realtime Gateway Trio | 48K lines of Go: three WebSocket gateways over Kafka + Cassandra, connection/session/presence separated. | app/gateways/ |
| 15 | 68 | Cloudflare Edge Platform | 10 Workers over D1, R2, KV, Durable Objects, Hyperdrive — including a WASM FFT fingerprinting DO. | app/workers/ |
| 16 | 65 | Solomon — Evaluation Gates & Promotion | Nine gates that can block a release, two of them latency and throughput; full artifact lineage to the running image. | Solomon/solomon/evaluation |
| 17 | 63 | Solomon — Training & Experiment Tracking | Noam + label smoothing, checkpoint/resume, MLflow, SHA256 run manifests. Val PPL 40.0 in 3 h on CPU. | Solomon/solomon/training |
| 18 | 62 | Dating Swipe Deck & Discovery UI | The largest mode in the app — and where the phantom-partition bug became visible. 49 audited findings with file:line. | IOS/TheOne/Dating |
| 19 | 61 | BoxOffice & the Moderator Console | "27 endpoints and curl" → a four-route operator console where a 502 on takedown is reported as a result, not a failure. | IOS/.../BoxOffice, /support/staff |
Tier C — breadth and depth-of-craft. Mention, don't lead. (<60)
| # | Score | Project | One line | Where |
|---|---|---|---|---|
| 20 | 57 | OpenHouse Global | Real-estate service with 7 versioned forward-only CQL migrations and a buyer journey modeled as a state machine. | app/OpenHouseGlobal |
| 21 | 56 | Billboard Player — Audio UI & Engine | Full music player: AVPlayer behind an actor, MTAudioProcessingTap, live spectrum, on-device vDSP FFT fingerprinting. | IOS/TheOne/Billboard |
| 22 | 55 | Palantir J8 — Exhibit Factory | Ontology-driven financial pipeline; data-driven normalizer registry; fail-closed ingest. | Palantir/J8 |
| 23 | 54 | Multiplatform Clients (umbrella) | 611K lines Swift / 13 modes, 19-module Compose port, Next.js platform. Decomposed into the five UI projects. | IOS/, Android/ |
| 24 | 53 | Solomon — The Transformer | Built from the Annotated Transformer, component by component, every deviation justified in writing. KV cache proven equal to full recompute. | Solomon/solomon/model |
| 25 | 51 | Web Platform UI | Every product mode in one Next.js app, compiled against the same protobufs the backend serves. Playwright + vitest. | app/web-platform |
| 26 | 46 | Solomon — Corpus, Tokenizer & Data Governance | 130M tokens cleaned and deduped, SentencePiece BPE (0 UNK), data card with contamination checks. | Solomon/solomon/data |
| 27 | 39 | Technical Documentation Corpus | Three 17–20 chapter manuals with gap registers — including a 14/52 Spotify parity analysis. | docs/ |
| 28 | 17 | Portfolio & Marketing Infrastructure | Cloudflare Pages site; a DNS migration plan whose step 2 is "verify the MX records or you lose email". | Portfolio/, Marketing/ |
Reading paths
Two umbrella entries decompose into the detail projects. They are listed once each and are not double-counted:
| Umbrella | Decomposes into |
|---|---|
| Solomon platform | Model · Corpus · Training · Eval gates · Serving/K8s · Retrieval/agent |
| Multiplatform clients | Billboard player · BoxOffice console · Dating swipe · Chat UI · Web platform |
If you have 20 minutes with a Google interviewer
1 → 3 → 23 → 28
If you have 20 minutes with Spotify
1 → 14 → 26 → 5 (the parity analysis)
If you have 20 minutes with a bank
2 → 12 → 22 → 11 (the restore-verify timer)
The five user interfaces, as a set
Billboard player · BoxOffice + moderator console · Dating swipe deck · GlobalChat messaging · Web platform
The six Solomon projects, as a set
Model → Corpus → Training → Eval gates → Serving/K8s → Retrieval/agent
Supporting documents
- RESUME-BULLETS.md — every project reduced to resume lines, pre-sorted by employer.
- INTERVIEW-STORIES.md — the eight stories worth telling, in situation/action/result form, with the follow-ups each invites.
Infographics
Rendered, self-contained HTML in ../infographics/:
| File | What it shows |
|---|---|
01-portfolio-ranking.html | All 28 projects as a scored tier board |
02-activity-graph.html | The outbox → Kafka → three-consumer pipeline, drawn |
03-platform-map.html | Every service in /srv and who calls whom |
04-scale-dashboard.html | The numbers: LOC, services, datastores, tests |
05-skills-matrix.html | Skills mapped to Google / Spotify / bank requirements |
06-solomon-pipeline.html | Corpus → training → eval gates → promotion → k8s |
The honest summary
/srv holds ~4.3M lines of authored code across 13 backend services, 10
edge Workers, 3 realtime gateways, 4 datastores, and 3 client platforms, plus
a from-scratch LLM platform and three book-length engineering manuals.
For the target role, three things carry the interview and the other twenty-five are supporting evidence:
- You built a canonical event-driven fan-out pipeline — outbox, log, consumer groups, idempotency — and can defend every trade-off in it.
- You found and fixed real distributed-systems bugs — a phantom Cassandra partition, a leaking Redis consumer-group PEL, a fatal map race under concurrent calls — and wrote each one down.
- You shipped and measured it, with runbooks, contract tests in CI, and named failure modes.
That is the profile. The rest is range — and there is a great deal of range.