About

Mohammed Faisal Parvez. AI Safety & Machine Learning Engineer: interpretability, LLM/VLM inference, model auditing. Hyderabad, India, open to relocation.

Two published AI-safety audits (Apart Research). Lead engineer on an autonomous robot's edge-ML stack. Author of independent reproducible benchmarks that publish their negative results. Python, Rust, PyTorch, from A100 cloud to 1-bit edge.

Builds systems that decide when a signal is trustworthy enough to act on. The probe that hit AUROC 1.00 and was refused; the confabulation floor that measures how much of a self-report is noise; the seven QuantBench configurations you cannot rely on; a controller that stages and guards before an irreversible kill; a citation gate before an answer; a gate before a parse.

Résumé (PDF)

Experience

Cybertronix, Hyderabad. Machine Learning & Robotics Engineer

Early-stage autonomous indoor-cleaning-robot startup.

  • Leads a 5-person perception and edge-ML team in production, owning the pipeline from A100 cloud training to on-robot ARM/Jetson deployment.
  • Orchestrates and benchmarks LLM/VLM inference across A100 cloud and ARM/Jetson edge for latency, memory and accuracy.
  • Built edge_monitor, a Rust safety-gated autonomous controller for GPU-bound AI workloads: per-process telemetry for co-located services (Ollama, vLLM, ROS2), plus an autonomous mode with VRAM and thermal policies and staged escalation, SIGTERM then SIGKILL with PID guards, before anything irreversible.
  • Published GPTQ- and AWQ-quantized LLMs (Qwen2.5-1.5B) to Hugging Face.
  • Diagnosed and fixed a 1-bit vision-grafting training failure: ~3× CPU speedup and ~6× lower memory than 4-bit, across a 6-pipeline VLM benchmark (Qwen2.5-VL, InternVL2).

Apart Research, research sprints

Secret Loyalties sprint (with Formation Research); Digital Minds sprint.

  • Detection ≠ Loyalty: a probe at AUROC 1.00 shown to be detecting fine-tuning drift, published as a false-positive mode for deceptive-alignment detection.
  • The Control You Cannot Run: the confabulation floor for introspective self-report, with pre-registered detection tests and their failures published.

Capabilities

Languages
Python, Rust, C/C++, JavaScript/TypeScript, SQL, Bash.
Methods
RAG, embeddings, structured output, LLM/VLM inference and deployment, white-box probing, model auditing, interpretability, evaluations, red-teaming and robustness.
Optimization
Quantization (GPTQ, AWQ, INT8, FP16, 1-bit/BitNet), inference optimization, cloud-to-edge deployment.
Frameworks
PyTorch, HF Transformers, vLLM, Ollama, llama.cpp, bitnet.cpp, ONNX, TensorRT, WebLLM/WebGPU.
Systems
FastAPI, PostgreSQL/SQLite, ChromaDB and vector DBs, Cloudflare Workers/D1, Docker, Modal, pytest/Playwright.

Education and speaking

  • B.E. Artificial Intelligence & Machine Learning, 2025. Lords Institute of Engineering & Technology, Osmania University, Hyderabad.
  • NPTEL: Responsible & Safe AI Systems (IIIT-H / IIT-M).
  • Research sprint talk, 2026. Presented the secret-loyalty result and its false-positive mode.
  • Speaker, Microsoft Student Club, 2024. Introduction to Azure, 150+ attendees.