Passionate AI Engineer | Research-Driven Developer | Enterprise Solution Architect
Building for fun and for impact
Hey, I'm Artur, and I'm obsessed with AI and software. To me, it's fascinating how so many of the world's problems can be solved with just 2 hands, a keyboard, and practically no labor. This obsession has allowed me to build stuff I never could've imagined. Just a few months ago I developed novel attention mechanisms for spatiotemporal understanding that hit 99% accuracy in context classification, and a year ago I built a neural network for chip power consumption prediction that surpassed Synopsys's benchmark! I've done plenty more projects but decided to highlight these gems. Also, I speak 5 languages which I think is pretty dope :)
Current Mission: At Deloitte, I'm building next-generation agentic RAG pipelines that autonomously discover industrial threats by fusing vulnerability data with stakeholder evidence.
- Natural Language Processing 📝 - Conversational AI and text analysis solutions
- Computer Vision 👁️ - Advanced image processing and visual understanding systems
- AI Agents 🤖 - Autonomous systems with decision-making capabilities
- High-Performance Computing ⚡ - GPU-optimized algorithms for large-scale AI workloads
I'm currently working on spatiotemporal news analysis, with my latest preprint available here. This research includes the SpatioTemporal News Corpus dataset and Space-Time-MiniLM model, both available on Hugging Face.
🧠 AI Mental Health Assistant for Lebanon - Culturally-Aware Healthcare AI
- 🏆 Built in 12 hours during an intense hackathon using the newly-released Llama 3
- 🎯 Localized for Lebanese context with insights from regional mental health professionals
- 🗣️ Multimodal interaction supporting both voice and text communication
- 🤖 Agent-based architecture using CrewAI for specialized assessment tasks
User Input (Voice/Text) → Speech Recognition → Agent Analysis → Follow-up Generation → Assessment Classification → Localized Resources
Key Innovation: Dual-agent system with specialized roles:
- Follow-Up Questioner: Analyzes responses and generates contextual follow-up questions
- Mental Health Classifier: Provides supportive assessments with Lebanese mental health resources
Technologies: Llama 3, CrewAI, Python, Speech Recognition, TTS Integration
⚡ SpMSpM: High-Performance Sparse Matrix GPU Kernels - Optimized ML Computing
- 🔥 190x speedup over CPU implementation on large datasets
- 📊 Systematic optimization through 10+ kernel iterations
- 🎯 Real-world application in pruned neural network layers
| Optimization Strategy | Impact | Implementation |
|---|---|---|
| Shared-Memory Tiling | Reduced global memory traffic | Per-block accumulation buffers |
| Privatization | Eliminated atomic contention | Block-local counters |
| Memory Coalescing | Maximized bandwidth utilization | Strided access patterns |
Performance Results:
Dataset Size: Small → Medium → Large → X-Large
CPU Baseline: 8.9ms → 33.3ms → 66.6ms → 914.2ms
Our GPU Kernel: 0.15ms → 0.42ms → 0.66ms → 4.79ms
Speedup: 59x → 79x → 101x → 191x
Applications: Efficient inference for pruned neural networks in resource-constrained environments
📡 RUDP: Reliable UDP Protocol Implementation - Network Systems Engineering
- 🔄 Full TCP-like reliability over UDP transport layer
- 📦 Custom 16-byte header with sequence numbers and control flags
- 🛡️ Comprehensive error handling with checksum validation and retransmission
- 📊 Network simulation with configurable loss rates and delays
Connection Management: 3-way handshake → Data Transfer → 4-way termination
Reliability Layer: Sequence tracking → ACK processing → ARQ retransmission
Error Recovery: Checksum validation → Fragment reassembly → State management
Key Features:
- Automatic Repeat Request (ARQ) with timeout-based retransmission
- In-order delivery guarantee through sequence validation
- Fragment storage system for out-of-order packet handling
- Real-time network statistics and performance monitoring
Impact: Demonstrates deep understanding of network protocols and distributed systems design
I'm always excited to collaborate on innovative AI projects, especially those that:
- Push research boundaries
- Create positive social impact
- Are crazy difficult
🔄 This README is updated regularly to reflect my latest projects and achievements. Last updated: 8/28/2025