Repositories

Open-source multi-modal fine-tuning. LLMs and vision models, fully local, accelerate edge deployments. All-in-one framework for fine-tuning large language models and vision models on your machine. Session-based training for Qwen3.8, GLM-5.3-Flash, Kimi K3, MiniMax-H3, DeepSeek-V4, Gemma 4 with LoRA/QLoRA. Plus vision models (ResNet, YOLO, CLIP). Privacy-first, runs completely local, deploy to edge devices. No cloud required.

Shipped: A 0→1 full-stack AI vision platform crushing real-time object detection at scale. We built Sentinel Cloud Vision from scratch with battle-tested ML frameworks (JAX, PyTorch, TensorFlow) powering insane inference speeds. FastAPI + PostgreSQL handle production workloads. Kafka streams real-time events, Cassandra captures metrics, Spark crunches data. 20+ shipped features: instant image upload, live camera streaming, custom model loading, data labeling, performance dashboards. Enterprise-grade microservices: isolated user/admin domains, zero-trust auth, comprehensive testing. 80+ docs prove it. Kubernetes-ready. Open-sourced for builders who want to skip 0→1 on vision AI.

Multi-stage deep learning framework for temporal emotion recognition in conversations. Tracks emotional shifts across dialogue sequences using visual, audio, and text modalities with Recurrent Memory Transformer. Modular codebase with clean package structure, YAML-based config management, and configurable paths. Stage 0 (preprocessing) extracts features from video. Stage 1 (multimodal fusion) combines frozen ResNet-50 + Wav2Vec2 backbones. Stage 2 (temporal modeling) processes utterance sequences with RMT for per-speaker trajectory tracking. Production-ready with reproducible results tracking via date-based naming conventions.