A complete guide to local AI in Swift: a URLSession-based OllamaClient with non-streaming and streaming chat using AsyncStream, SwiftUI integration with async token display, configuring Ollama for LAN access from iOS devices on the same network, and Apple Foundation Models for true on-device inference on iOS 18+ and macOS 15+ using the Neural Engine with zero network access required.
A practical guide to PyTorch Automatic Mixed Precision for ML engineers: the numerical difference between fp16 and bf16 and when to use each, complete AMP training loop with autocast and GradScaler, how GradScaler adaptive scaling works and how to tune it, which ops autocast converts vs keeps in float32, AMP with HuggingFace Trainer, mixed precision inference with permanently converted model weights, and a hook-based debugger for finding which operation first produces NaN under fp16.
A complete guide to Ollama integration in C# and .NET: the OllamaSharp library for chat and model listing, Microsoft.Extensions.AI with OllamaChatClient registered as IChatClient for dependency injection, streaming with IAsyncEnumerable, Semantic Kernel with AddOllamaChatCompletion and prompt templates, an ASP.NET Core streaming SSE endpoint, and generating embeddings with cosine similarity in C#.
A practical guide to Online Hard Example Mining (OHEM) for ML engineers: how the forward-select-backward mechanism concentrates gradient signal on informative samples, PyTorch implementation with per-sample cross-entropy and topk selection, segmentation-specific OHEM with threshold-based pixel selection, OHEM vs focal loss and when to combine them, OHEM for metric learning and embedding training, and tuning the keep_ratio hyperparameter without destabilising training.
A complete guide to integrating Ollama in Java Spring Boot applications: adding spring-ai-ollama-spring-boot-starter, configuring base URL and model in application.yml, chat with ChatClient including streaming with Flux, prompt templates with system prompts and parameter injection, generating embeddings with EmbeddingModel and computing cosine similarity, and a direct HTTP approach using RestClient with record-based request/response types for full control without Spring AI.
A practical guide to Matryoshka Representation Learning (MRL) for ML engineers: how nested dimension training works, fine-tuning MRL models with sentence-transformers MatryoshkaLoss, using truncated embeddings at inference time with correct renormalisation, building two-stage RAG retrieval with a small-dimension FAISS index for recall and full-dimension reranking for precision, benchmarking quality across dimensions on your own domain data, and how MRL compares to PCA-based dimensionality reduction.
A practical guide to monitoring Ollama with Prometheus and Grafana: what performance metadata the Ollama API exposes in every response, a Python Prometheus exporter that scrapes ollama ps and api tags, a FastAPI middleware proxy that intercepts requests and records request count, latency histograms, and token counters, Prometheus scrape configuration, and key Grafana dashboard PromQL queries for models loaded, VRAM usage, request rate, tokens per second, and P95 latency.
A practical guide to pytest for ML engineers: structuring test suites by speed and scope, shared fixtures for tiny models and small batches, testing data preprocessing deterministically, model shape and gradient flow tests, the overfit test for catching silent training bugs, loss function correctness tests, and configuring pytest markers and GitHub Actions CI to run fast unit tests on every push and GPU integration tests on a schedule.
A complete guide to building an Ollama-powered Telegram bot with python-telegram-bot: a basic bot that responds to messages with typing indicator, per-user conversation history with trimming, /start, /clear, and /model commands, restricting access to specific Telegram user IDs, and running the bot as a persistent systemd service alongside Ollama.
A practical guide to W&B Sweeps for ML engineers: how the sweep controller and agent architecture works, configuring Bayesian vs random vs grid search with the right parameter distributions, writing sweep-compatible training scripts that read from wandb.config, running parallel agents across multiple GPUs and SLURM clusters, using Hyperband early termination to save compute, interpreting parallel coordinates plots, and avoiding common pitfalls like over-broad search spaces and treating the best sweep run as a final result without seed averaging.
