Debnil Dey

AI Engineer · Multi-Agent Systems · Computer Vision & CUDA · MLOps

⚡ Systems & AI Engineering Avatar

🎨 Dream • Focus • Create • Inspire • Repeat

debnil= {
"name": "Debnil Dey",
"role": "AI Engineer & Systems Builder 🤖",
"location": "📍 Kolkata, West Bengal, India",
"specializations": [
"Autonomous Multi-Agent Systems (ADK, MARL)",
"Real-Time Computer Vision & TensorRT/CUDA",
"Production MLOps & Real-Time Drift Engineering",
"Multimodal GenAI & Model Context Protocol (MCP)",
"3D WebGL & Spatial Digital Twins",
],
"current_stack": ["PyTorch", "FastAPI", "Next.js", "Gemini 2.5/3.5 Flash", "Three.js", "Docker"],
"fun_fact": "Built real-time AR gesture engines with CUDA & WebGL GLSL shaders ⚡",
}

• 🔭 Passionate about engineering scalable, production-grade AI systems combining deep learning, real-time spatial computing, and agentic workflows.
• 🌌 Multi-Agent Systems & MARL: Built real-time 50Hz sports digital twins with PyTorch MARL gRPC inference and supply chain RL optimizers yielding 36% cost savings.
• 👁️ Computer Vision & AR: Developed TensorRT FP16 / CUDA accelerated gesture engines (Sanctum) & ONNX spatial perception pipelines with Grad-CAM SHAP explainability (ChronoSpatial Engine).
• 🛡️ MLOps & Observability: Engineered multi-statistic real-time drift detection (PSI, KS-Test, EWMA, STL decomposition) in Sentinel and autonomous self-healing agent swarms with Google ADK in Aegis.
• 💬 Ask me about Multi-Agent Systems · Computer Vision & CUDA · MLOps · WebGL · GenAI

Distributed Multi-Agent Tactical Synthesizer & 3D Digital Twin for FIFA World Cup 2026

Multi-Agent Self-Healing Infrastructure & Incident Remediation Platform

• 🤖 4 Cooperating AI Agents built on Google Agent Development Kit (ADK) powered by Gemini 2.5 Flash:
  • 🛡️ Sentry: Telemetry baseline deviation & anomaly tracking.
  • 🔍 Sleuth: Multi-service log correlation & ChromaDB vector memory retrieval.
  • 💊 Medic: Risk-tiered remediation proposal with human-in-the-loop approval gates.
  • ✍️ Scribe: Automated markdown postmortem document generation.
• 🧪 Failure Injection Simulator: Real-time chaos engineering harness for model drift, GPU memory leaks, and pipeline bottlenecks.

Low-Latency Real-Time AR VFX Engine accelerated with PyTorch, TensorRT FP16 & CUDA

• 🤲 MediaPipe Hand Tracking$\rightarrow$ Scale-invariant 21-point wrist-relative normalization $\rightarrow$ 30x63 sliding window tensor.
• ⚡ GPU Acceleration: PyTorch 1D-CNN compiled into TensorRT FP16 engines & custom CUDA particle physics kernels.
• 🎨 GLSL Shaders: Multi-octave Fractal Brownian Motion (FBM) portal shaders & screen-space optical flow reality warping in Three.js.

Real-Time Perception, Object Tracking & Explainable Spatial-Temporal Risk Assessment

• 🎯 Object Perception: YOLOv8 / MobileNetV3 backbone paired with perspective velocity and distance estimation algorithms.
• ⚡ Quantized Inference Engine: MobileNetV3 128-d spatial embeddings fused with 15-d track motion vectors into ONNX Runtime quantized ANN.
• 🔬 Auditable Explainability: Integrated Grad-CAM spatial heatmaps and Integrated Gradients SHAP temporal attributions.

Browser-Native AR Hand Tracking & Electromagnetic Physics Simulation

• 🤲 Dual-hand 21-landmark tracking via MediaPipe Vision WASM (GPU accelerated).
• 🧲 Gesture EM Force Model: FIST (100% pull force), PINCH (35% precision force), OPEN (gravity release).
• ⚡ WebGL 2.0 shaders, 200-particle systems, Cannon-es physics, live spatial geometry theorems.

Real-Time Data Drift Engine, Probability Calibration & Explainability Stack

Clinical Decision Support System with SHAP & UMAP Explainability

Multi-Agent Supply Chain Optimization with Reinforcement Learning

• 💰 36% Cost Savings compared to traditional heuristic baselines.
• 🛡️ 0.998 / 1.000 Resilience Score under simulated supply chain disruptions.
• 📦 99.7% Order Fill Rate trained across 913K rows of multi-store demand history.

Digital Twin Carbon Sandbox with Gemini 2.5 Flash Receipt OCR & Streaming EcoCoach

• 🤖 Powered by Gemini 2.5 Flash for multi-modal structured receipt scanning OCR.
• 💬 Live streaming Server-Sent Events (SSE) AI EcoCoach advisor.
• 🎨 Framer Motion micro-animations & Zustand local state persistence.

Multilingual GenAI Operations Layer for Fans, Staff & Organizers

• 💬 6 Integrated GenAI Modules: Multilingual stadium chat (12 languages), Crowd Intelligence advisory, Accessibility plain-language simplify with TTS, Eco-route finder, Interactive SVG map, and Volunteer briefing.
• ⚡ Powered by Google Gemini 3.5 Flash with reasoning tokens, hosted on Vercel Edge.

Voice-Enabled AI Assistant with Iron Man HUD Aesthetics

• 🎙️ Real-time voice I/O via Web Speech API paired with streamed Gemini 2.5 Flash responses.
• ⚡ Animated Arc Reactor HUD, particle canvas, system diagnostics, and career tracking panels.

Local Technical Debt & Sprint Planning Model Context Protocol (MCP) Server

• 🔍 Scans repository ASTs for unresolved TODO:, FIXME:, BUG: tags (scan_todos).
• 📜 Dynamically constructs read-only project://backlog virtual markdown resources for LLMs.
• 🎯 Includes sprint-planner prompt template for automated AI sprint breakdown.

🤖 AI, Deep Learning & MAS

👁️ Computer Vision, CUDA & WebGL

🖥️ Backend & MLOps Infrastructure

🌐 Frontend & Web Engineering