Mohammed Faisal Parvez.
I study how vision-language models fail — and make them run where they shouldn't.
Research interests
VLM reliability & calibration · extreme quantization (1-bit) ·
edge inference · sim-to-real transfer
Scroll to explore
Failure is data_
6Pipelines benchmarked on A100s
~2× / 5–10×Precision / recall via selective refusal
~3× · 6×CPU speedup · less memory @ 1-bit
0.676Real-world mIoU · indoor seg
1,458Images hand-labeled
97.3%CIFAR-10 ensemble accuracy
01/
About
Independent researcher in Hyderabad studying where modern perception stacks break: calibration under uncertainty, benchmark contamination, extreme quantization, and the sim-to-real gap.
By day — lead of a 3-person perception & edge-ML team at Cybertronix, owning the full stack from cloud A100 benchmarking down to ARM deployment on an autonomous indoor robot. Full bio →
02/
Featured Research
F-01VLM Calibration · Modal A100
Selective Refusal as Protective Calibration
Which VLM should a safety-critical robot trust when it's uncertain?
~2× precision · 5–10× recall Read finding → F-031-bit VLMs · EdgeVision on 1-bit LLMs & a Silent Training Failure
Can vision be grafted onto extreme-quantized (1-bit) language models?
~3× CPU speedup · ~6× less memory Read finding →