Pinned Loading

1. llm-alignment-pipeline llm-alignment-pipeline Public

   Fine-tune Mistral-7B with QLoRA, align with Direct Preference Optimisation (DPO), and serve with vLLM — a complete, production-quality LLM alignment pipeline running on a single A100 GPU.

   Python

2. ml-serving-platform ml-serving-platform Public

   Production MLOps platform with FastAPI model serving, PSI-based drift detection, automated retraining, A/B traffic splitting, and Prometheus+Grafana observability — deployable via Docker Compose or…

   Python

3. aerial-sematic-segmentation aerial-sematic-segmentation Public

   Custom U-Net with EfficientNet-B4 encoder for aerial imagery segmentation. CutMix augmentation, boundary-aware loss, ONNX export, FastAPI serving. 52.1 mIoU on LoveDA (+6.4 over baseline).

   Python

4. sae-interp-dashboard sae-interp-dashboard Public

   From-scratch Sparse Autoencoder training on GPT-2 Small activations + interactive Next.js dashboard for exploring learned features. Built for mechanistic interpretability research and portfolios.

   Python

5. explain-this-model explain-this-model Public

   Upload any HuggingFace transformer model + prompt — get ranked, visualised neuron/attention head activations with natural-language explanations via Groq, Gemini, or Claude. Model-agnostic, CPU-only…

   Python

6. circuit-discovery-toy-task circuit-discovery-toy-task Public

   Reverse-engineering a grokked transformer on modular addition: Fourier mechanisms, activation patching, and minimal circuit identification. A complete mechanistic interpretability case study built …

   Python