Blog
Technical notes, implementation write-ups, and research reflections in MDX.
Optimus: Autonomous EnergyPlus Building Controller
A safety-constrained physical-AI controller for autonomous building operations. Uses EnergyPlus simulation with PyEnergyPlus API, local LLM (Ollama/Qwen) or deterministic fallback for bounded planning, and hard safety validation before schedule override injection.
Why Distributed Training Still Bottlenecks at Scale
A practical look at gradient synchronization, communication overlap, and the hidden costs of scaling PyTorch DDP. Based on building a distributed ML training framework for heterogeneous consumer hardware.
Technical Note: Designing Reliable Multi-Agent Systems
Notes on orchestration, context management, and fault tolerance for specialized LLM agent teams. Based on building AgentForge and AI Research Orchestrator.
AgentForge: Multi-Agent Software Engineering with Local LLMs
Building a local-first multi-agent coding assistant where specialized agents (Analyst, Architect, Planner, Critic, Consensus) collaborate on software tasks—running entirely on your machine with Ollama. Focus on role specialization, structured intermediates, and local execution.
Building Real-Time Speech Emotion Recognition
How I built a complete speech emotion recognition pipeline from audio preprocessing to real-time inference—published at IEEE DECoN 2025. Covers the published lightweight CNN approach and the extended project implementation with CNN-BiLSTM.
AI Research Orchestrator: Multi-Agent Literature Synthesis
An autonomous research assistant that plans literature reviews, retrieves papers from academic APIs, analyzes findings, and generates structured reports—built as a modular multi-agent system in TypeScript with production and offline modes.
Brain MRI Tumor Segmentation: 3D U-Net to Attention + EfficientNet
An iterative research journey through volumetric brain tumor segmentation—comparing baseline 3D U-Net, EfficientNet encoder, and Attention U-Net on the BraTS dataset with patch-based training and mixed precision. Final validation Dice: 0.8256 with Attention + EfficientNet.
Weather Prediction at Scale: 4,300+ Cities, 90GB, One Model Each
How I built a city-specific weather forecasting pipeline training 4,300+ independent XGBoost models on 10 years of historical data—lessons in data organization, pipeline automation, and when local models beat global ones.