shreyansh

12 Interviews Resume Positioning

Part of Index of Inference Engineering. Book: Ch 0 framing + Appendix B (Recommended Reading). Target: $75-95k to net 5L/mo after EOR fees/FX. See tool repo profile.yaml for filters.

Position (headline)

Backend / Platform Engineer (Java, K8s) + LLM Inference (vLLM) — rare combo. Java+K8s is the moat; most ML devs lack prod K8s.

Resume bullets (impact + remote signals)

  • "Own vLLM on K8s-GPU (L4): p95 TTFT X, Y tok/s, preemptions 0, $/MTok Z — Helm+KEDA queue-driven autoscale."
  • "Quant parity: AWQ-INT4 saved N GB at ≤1% qualityΔ; router holds SLO while minimizing $."
  • "Spring gateway: auth/ratelimit/Kafka→AsyncLLMEngine, per-tenant budgets, audit."
  • Remote signals every bullet: async ownership, SLOs, English docs/dashboards, overlap IST→EU/US.

Where to apply (only offshore-friendly)

Worldwide/async (e.g. Inferact-type 200−400kworldwide),EMEA+APAC/India−explicit(IST3.5hCEToverlaphireable),contractor−via−EOR(200-400k worldwide), EMEA+APAC/India-explicit (IST 3.5h CET overlap hireable), contractor-via-EOR (40-80/hr = 6-13L/mo, easiest to 5L: Arc/Turing/Lemon/Toptal/Deel), YC/Wellfound Remote+EOR. Skip LinkedIn Remote US (90% US-only noise) and UTC-1..+3 hard posts.

Interview prep (2x/wk)

  • Whiteboard: request path → scheduler → KV check → prefill/decode → metrics → $.
  • Tune live: given p95 spike + num_requests_waiting climb, answer "shorten max-model-len → lower max-num-seqs → raise utilization" + why.
  • Code: FastAPI streaming proxy, KEDA ScaledObject YAML, Spring rate-limit filter, parity harness.
  • System design: reuse Index of High Level Design + add GPU chapter (this track).

Closer

1 small PR to vLLM / production-stack / llm-d (docs/fix + benchmark). Cuts through "no prod inference exp" objection.

Talk track + reading (Book Ch 0 / App B)

  • Open with the three-layer frame: "Inference engineering spans runtime, infrastructure, and tooling, from CUDA to Kubernetes." Then place each project on a layer — shows systems thinking, not just tool usage.
  • Emphasize open vs closed: you can tune latency/cost/control on open weights in ways an API can't offer (00b).
  • If asked where to go deeper, name a paper from Appendix B by topic (architecture, GPU infra, optimization research) — reading-list fluency signals depth. See 13-Resources-Glossary.