Arjun spent nine years inside Indian AI infrastructure — model-serving at a fintech, then GPU benchmarking for a hardware-review publication. His focus is the small, important corner of consumer ML where a published tokens-per-second number is reproducible by a reader on the same hardware.
Background
Arjun trained as a computer scientist at IIT Madras and joined an Indian fintech to work on model-serving infrastructure before moving to GPU benchmarking full-time at a major hardware-review outlet. His writing focuses on the question that matters most to consumer reviewers in 2026: which open-weight model actually runs at the speed the marketing says, on the hardware the reader can afford.
Vague impressions of cloud-API speed are unhelpful — local AI benchmarking is the first time most consumer reviewers can publish numbers that are reproducible by their readers. Arjun Mehta, on reproducible AI benchmarks
Career timeline
ML serving engineer, Bangalore fintech
Built and maintained model-serving infrastructure for production credit-scoring and fraud-detection systems.
Senior benchmarks engineer, hardware-review publication
Designed reproducible benchmark suites for consumer GPUs and Apple Silicon — CUDA, ROCm, Metal.
ML Performance Engineer, LM Studio
Long-form coverage of local-LLM throughput, quantization formats and consumer-AI tooling for Mac and Windows.
Editorial principles
Every benchmark published here is run on a sterile machine with the same prompt set, the same model file and the same SDK version across hardware. Numbers are measured, not estimated. Disclosures appear only at the end of a piece and never influence rankings.
Contact
Arjun Mehta reads every email but cannot offer one-to-one support for the LM Studio application itself — for that, please use the publisher's official Discord and documentation. For corrections, story tips or speaking enquiries, reach out via the address listed on the main site.
← Back to the LM Studio guide