qaisermehdi
Remote: Yes
Willing to relocate: No
Technologies: vLLM, SGLang, Prometheus /metrics, Python, GPU benchmarking (T4, L4, L40S, RTX 3090, A100)
Résumé/CV: https://github.com/qaisermehdi3-coder
Email: [email protected]
Looking for fixed-price contract work: I measure what self-hosted LLM inference really costs. GPU cost per request at your real load, how much paid GPU time serves nothing, and which GPU is cheapest for your answer-time target. I read only vLLM's or SGLang's own /metrics, so nothing in your stack changes. If you won't run my script on your servers, I rent the same GPU, run your model and settings, and send the report.
Sample (RTX 3090 vs A100, Qwen 7B, vLLM 0.27.1): at the same load the 3090 was cheaper per token; at a 6-7 s answer time the A100 was about 4.6x cheaper. https://github.com/qaisermehdi3-coder/qvunex/tree/main/bench...
Try it: which measured GPU is cheapest at your answer time, one file, standard library: https://github.com/qaisermehdi3-coder/qvunex/tree/main/bench...
Recent public work: pydantic-ai fixed a $0-cost bug I reported (#9491); AgentMeasure runs my self-hosted evidence case in their CI; sent fixes for Kedify's vLLM autoscaling examples (one approved); measured two vLLM servers on one GPU for vLLM #60599.
Free first step: send your model, vLLM flags and typical prompt/answer length, and I'll run a short mini-report on a rented GPU. The full report is $1,500 for one setup, $3,000 for a fleet, about a week, half upfront.