Mohammed Faisal Parvez

I lead Cybertronix's perception and edge-ML pipeline, taking models from A100 cloud training to ARM and Jetson robots in production. I also benchmark LLM/VLM inference for latency, memory, and accuracy across cloud and edge environments.

I built edge_monitor in Rust to govern GPU-bound AI workloads, with telemetry for Ollama, vLLM, and ROS2 plus safety-gated VRAM and thermal policies. I've published GPTQ/AWQ-quantized Qwen2.5 models and resolved a 1-bit vision-grafting training failure across a six-pipeline VLM benchmark.

My AI-safety research challenges false positives in deceptive-alignment detection: my relative spectral probes and valence-gap diagnostic distinguish hidden loyalties from ordinary fine-tuning drift. I released a reproducible audit harness and presented the result in a research sprint talk.

I also build practical, auditable AI products, including Meritline, a deterministic CV-shortlisting platform, and Filing Reader, a privacy-first RAG system for financial filings. My work centers on reproducibility, robust evaluation, and safety from A100 cloud infrastructure to 1-bit edge deployment.