CV
Contact Information
| Name | Muhammad Anas Raza |
| Professional Title | Ph.D. Candidate & ML Systems / Research Engineer |
| memanasraza@gmail.com | |
| Location | Rochester, Michigan 48326 |
Professional Summary
Machine Learning Systems and AI researcher with 5+ years of combined research and industry experience. Author of 12 papers (7 first-author, 230+ citations), recipient of the IEEE IJCB 2024 Best Poster Award, and open-source contributor with 20+ merged pull requests to Keras 3. Leading development of timm.cpp (zero-overhead C++ vision model inference engine) and multi-backend machine learning frameworks (K3Node, K3IM, Keras-Climate).
Experience
-
2022 - Present Rochester, MI
Graduate Research & Teaching Assistant
Oakland University
Conduct advanced research in ML systems, agentic AI, and software reliability, while mentoring students and assisting graduate courses.
- ML Systems & C++ Inference: Spearheading timm.cpp, a zero-Python C++17 inference engine running 106 PyTorch/timm architectures via ggml/GGUF, achieving up to 25.8x faster cold-start latency via zero-copy POSIX mmap and 3.4x-5.8x weight reduction with calibrated quantization.
- Multi-Backend Frameworks & Keras 3: Authored 20+ merged PRs to core Keras 3; engineered K3Node (multi-backend GNN library with 100% PyG public API parity across PyTorch, TensorFlow, and JAX with 700+ unit tests per backend) and K3IM (50+ vision and spatiotemporal model architectures).
- Agentic AI & Neural Architecture Search: Designed AutoMAS (end-to-end multi-agent software generation framework across 5 cooperative layers, achieving 82.75% Pass@1 on real-world benchmarks) and NeuroGenesis (reasoning-guided multimodal NAS across audio, video, image, and time-series data).
- Deep Learning Reliability & Fault Localization: Developed LocatorGraph (GNN-based bug localization on DL computation graphs, published in Empirical Software Engineering) and co-authored SelfHeal and RAGFix for automated program repair.
- Multimodal Forensics & AIOps: Researched deepfake detection on 100K+ videos across DFDC and FakeAVCeleb, winning the Best Poster Award at IEEE IJCB 2024 (RuleBoost); created DeepTrAnS for unsupervised server telemetry anomaly detection (99.88% precision, IEEE ICSC 2025).
- Teaching & Outreach: Graduate Teaching Assistant for Machine Learning and Software Testing; designed and delivered NAIRR Edge AI workshop modules.
-
2020 - 2021 Islamabad, Pakistan
AI Software Engineer
SenseDigital
Built industrial IoT pipelines, automated time-series analytics, and edge AI monitoring infrastructure.
- IoT Ingestion Pipeline: Architected automated BACnet data ingestion pipelines capturing building sensor telemetry for real-time HVAC efficiency analytics.
- Edge AI Deployment: Deployed real-time data processing, monitoring, and inference containers on NVIDIA Jetson Nano hardware.
- Telemetry Dashboards: Containerized InfluxDB time-series storage and Grafana visualization dashboards with Docker Compose on Linux.
Education
-
2022 - 2026 Rochester, MI
Ph.D.
Oakland University
Computer Science & Informatics
- Research Focus: ML Systems, Agentic AI, and Deep Learning Reliability
-
2022 - 2024 Rochester, MI
M.S.
Oakland University
Computer Science
- Visual Computing, Object Oriented Analysis, Advanced Data Structures & Algorithms
Honors and Awards
-
2024 Best Poster Award
Awarded for RuleBoost: A Neuro-Symbolic Framework for Robust Deepfake Detection.
Certifications & Personal Development
Skills
Languages
Projects
-
timm.cpp: Vision Model Inference in C++
High-performance C++ inference engine executing PyTorch/timm vision models via ggml/GGUF with zero Python runtime dependency. Hand-ported 106 vision architectures with strict numerical parity (Cosine Similarity >= 0.999997), up to 25.8x faster cold-start load times via zero-copy mmap, and 3.4x-5.8x calibrated quantization weight reduction.
- Stack: C++17, ggml, GGUF, PyTorch, CMake, Docker, CUDA, Metal, Vulkan
-
K3Node: Multi-Backend Graph Neural Networks
Next-generation GNN library with 100% public API parity with PyTorch Geometric (PyG), executable across TensorFlow, PyTorch, and JAX. Implemented 65+ conv layers, 26 aggregations, 31 pooling operators, 5 pre-trained Foundation Models (GraphMAE2, Graphormer, GraphGPS, GROVER, Mole-BERT), and 700+ unit tests per backend.
- Link: https://github.com/anas-rz/k3-node
-
K3IM: Keras 3 Image Models Zoo
Open-source deep learning library providing 50+ modern vision and spatiotemporal architectures across 1D, 2D, and 3D data modalities with cross-backend support on TensorFlow, PyTorch, and JAX. Features hand-implemented ViTs, Swin, ConvMixer, gMLP, and MLP-Mixer.
- Link: https://github.com/anas-rz/k3im
-
Keras-Climate: Earth Observation & Climate Modeling
Open-source multi-backend framework for remote sensing, climate forecasting, and Earth observation. Implemented 4-band UNet and SegFormer, PatchTST for multivariate forecasting, ConvLSTMNowcaster, and the NASA/IBM Prithvi geospatial foundation model.
- Link: https://github.com/anas-rz/keras-climate
-
AutoMAS: Automated Multi-Agent System Generation
Autonomous framework generating functional multi-agent software systems from natural language code intent across 5 cooperative layers. Evaluated on 91 real-world agentic systems with 82.75% Pass@1 and 89.78% accuracy. Published in IEEE QRS 2026.
-
NeuroGenesis: LLM-Driven Architecture Search
Reasoning-guided Neural Architecture Search (NAS) using LLMs to formulate and justify iterative model designs across audio, video, images, and multivariate time-series data with 20 candidate backbones. Published in IEEE QRS 2026.
-
LocatorGraph: GNN Fault Localization for Sequence Models
Novel fault localization framework converting sequence model code into TraceGraphs to localize deep learning bugs. Achieved 89.46% AUC on bug identification and 81.54% accuracy in pinpointing faulty nodes across 152 models. Published in Empirical Software Engineering (EMSE 2025).
-
RuleBoost: Neuro-Symbolic Deepfake Detection
Neuro-symbolic detector fusing deep perceptual representations (gMLP, ViT, TokenLearner) with geometric facial rules. Achieved 96.19% AUC on World Leaders Dataset and 95.44% on FaceForensics++. Awarded Best Poster Award at IEEE IJCB 2024.
-
DeepTrAnS: Unsupervised Server Telemetry Anomaly Detection
MLP-Mixer-based unsupervised anomaly detection framework for high-dimensional server telemetry. Achieved 99.88% precision, 99.73% recall, and 99.81% F1-score, significantly outperforming TranAD and OmniAnomaly. Published in IEEE ICSC 2025.
-
Real-Time Food Delivery Ecosystem & Apollo GraphQL Backend
Production-grade microservice and client ecosystem featuring an Apollo Server 4 GraphQL backend with real-time WebSocket subscriptions, React 19 web frontend, and cross-platform React Native / Expo SDK 53 customer and manager mobile apps.