ML Engineer | Computer Vision | Multimodal Retrieval | Architecture-Level R&D

ML Engineer focused on designing efficient, interpretable ML systems under real-world constraints.

• Build CPU-efficient pipelines with measurable latency–accuracy trade-offs
• Study model behavior (failure modes, drift, sensitivity) beyond benchmark scores
• Develop architecture-level modifications (not just training optimizations)
• Create modular, reproducible experiment systems for controlled evaluation
• Translate research ideas into deployable, constraint-aware systems

• Cross-version design (v8 → v11 hybrids) with backbone–head decoupling
• Custom modules integrated via shape-consistent YAML pipelines
• Focus on tiny-object sensitivity & feature preservation
• Profiling FLOPs, latency, and structural efficiency (CPU-bound)

• Implemented SORT / DeepSORT / OCSORT under unified pipelines
• Analyze ID-switches, track continuity, and failure cases
• Develop proxy metrics beyond MOTA for real-world reliability
• Visual diagnostics: trajectory behavior & temporal consistency

• CLIP-based embedding systems (image–text alignment)
• ANN search using NNDescent / FAISS
• UMAP clustering for structure discovery
• Metadata-aware filtered semantic retrieval

• Fully reproducible pipelines (config-driven experiments)
• Structured ablations for architecture validation
• Controlled studies on noise, resolution, and scaling effects
• CPU-first debugging for true bottleneck identification

• YOLOv11 + MOT benchmarking system
• Unified evaluation across SORT / DeepSORT / OCSORT
• Metrics: ID-switches, fragmentation, track lifetime
• Includes trajectory visualization & sequence diagnostics
• Built for real-world tracker comparison

🔗 https://github.com/Y-R-A-V-R-5/VisionTracker

• Cross-generation YOLO architecture experimentation
• Backbone–head hybridization via modular configs
• Focus: tiny-object detection & efficiency trade-offs
• Benchmarks FLOPs vs CPU latency vs accuracy
• Rapid testing of custom structural variants

🔗 https://github.com/Y-R-A-V-R-5/YOLO-Tweaks

• Custom backbone for plant disease detection
• Modules: SerpensELAN, DySnake, CBAM
• Optimized CPU-only inference pipeline
• Tuned for domain-specific feature patterns
• Evaluated under lighting & texture variability

• CLIP (ViT-B/32) embedding pipeline
• ANN indexing via NNDescent
• UMAP-based clustering & visualization
• Metadata-conditioned semantic search
• Designed for low-latency large-scale retrieval

ML & Vision: PyTorch, YOLO (v8–v11), OpenCV
Retrieval: CLIP, FAISS, NNDescent
Engineering: Python, Git, modular ML systems
Analysis: Matplotlib, Seaborn

• Long-context vision architectures
• Efficient MOT without heavy re-identification
• Multimodal fusion beyond embedding similarity
• CPU-first architectural optimization
• Model sensitivity to noise, scale, and distribution shift

• Ablation-first design to isolate architectural impact
• Track stability, variance, and failure patterns
• Ensure reproducibility across seeds & configs
• Prefer controlled experiments over large opaque runs
• Focus on understanding model behavior, not just outputs

GitHub: https://github.com/Y-R-A-V-R-5
LinkedIn: https://www.linkedin.com/in/yravr/
Email: adithyavardhanreddy2003@gmail.com