Back to Projects
★ FEATURED PROJECTDEEP LEARNING & NLP

AdaptML — Live Model Training & Inference Playground

Upload any dataset — CSV, PDF, images, or JSON. The system auto-detects the ML task, trains a model with a live loss curve, and lets you query it instantly with built-in explainability.

PythonPyTorchscikit-learnXGBoostHugging FaceLoRAChromaDBFastAPIWebSocketsReactSHAPGradCAMONNXDocker
Step 1 — Upload Your Data

Drag & drop any supported file or click to browse

Drop files here or click to upload

CSV · Excel · PDF · TXT · DOCX · JSON · JPG · PNG

Technical Architecture

Adaptive Pipeline Router

File type detection → automatic preprocessing pipeline selection. Tabular → pandas profiling + feature engineering + XGBoost/neural net. Text → chunking + embedding + RAG or LoRA fine-tune. Image → augmentation + transfer learning.

"Learns on the Spot" Engine

Small data: full training loop via scikit-learn. Text: RAG (fast mode) with ChromaDB, or LoRA fine-tune (deep mode) on Phi/TinyLlama. Image: frozen backbone + trainable classification head. All training streams results via WebSocket.

Explainability Layer

SHAP for tabular predictions. Attention heatmaps for transformer outputs. GradCAM for image classifications. Every prediction comes with a 'Why?' toggle.

Engineering Capabilities
  • End-to-end ML systems — not just notebooks
  • Dynamic routing between supervised, unsupervised, and deep learning
  • Fine-tunes LLMs/SLMs on custom data (LoRA, RAG)
  • Production UIs that anyone can interact with
  • Explainability and trust built into every prediction
  • Multi-modal data handling in a single system
  • Real-time training feedback with WebSocket streaming

Want a deeper walkthrough?

I can walk through the pipeline routing logic, model selection heuristics, real-time training architecture, and the explainability integration.