Senior ML Infra Engineer — Self-Host + Fine-Tune Open-Source Coding LLMs (SLM + 32B + 671B MoE)

30-Day Fixed-Price Project — Self-Host + Fine-Tune Open-Source Coding LLMs — $8,000 USD Fixed price: $8,000 USD. 30 calendar days. 4 payment milestones. Remote. Any time zone. We need a senior ML infra engineer to stand up an open-source coding LLM stack on our AWS GPUs in 30 days. This is a fixed-scope project with fixed price and hard deadlines. Not a retainer, not a discovery engagement. If the milestones below land, there is a clear follow-on. If they don’t, we part ways cleanly. The scope — exactly what you will deliver By day 30 we will have three open-source coding models serving traffic behind our gateway on our AWS GPUs, one fine-tuning pipeline running weekly, and a signed benchmark card comparing Pixis versus Anthropic Haiku 4 on a real coding workload. Deliverable 1 — Bucket S: pixis-code-3b in production Serve Qwen2.5-Coder-3B on 2× H100 with vLLM 0.7, FP8, speculative decoding. • p50 first-token under 80ms on FIM autocomplete • 200 requests/sec sustained per GPU • Wired behind our LiteLLM gateway with OpenAI-compatible endpoints • Basic Grafana dashboard: p50/p95/p99, tokens/sec, GPU utilization Deliverable 2 — Bucket M: Qwen3-Coder-32B in production Serve Qwen3-Coder-32B on 2× H200 with FP8 and prefill sharing. • p50 under 400ms for chat + edit at 60 req/s • Shadow-run 100% of our current chat traffic against Haiku 4 for 5 days • Produce a signed benchmark card: chat helpfulness win rate, cost per 1M tokens, p50/p95 latency Deliverable 3 — Weekly LoRA fine-tuning pipeline for Bucket S • Training script (PEFT or unsloth) that takes IDE accept/reject events and produces a LoRA adapter • Adapter hot-swap at inference (under 5 seconds, no server restart) • Eval gate that blocks bad adapters (baseline pass rate + accept-rate lift threshold) • First adapter trained on our first customer’s data, showing measured lift versus base Qwen2.5-Coder-3B Deliverable 4 — Runbook + handoff • Written runbook: how to deploy, how to roll back, how to retrain, common failure modes • 2-hour recorded walkthrough with our team • All code in our GitHub, all infra in our AWS account