Articles for people who
ship the stack
Rewrites on React, Node.js, TypeScript and AI. A translated body is shown when it exists. If it does not, you see the English text and a note that says so.
Tagged: LoRA
Training Enterprise AI Agents: Choosing Between SFT, DPO, GRPO and RAG
Training enterprise AI agents: when to use SFT, DPO, GRPO, LoRA or plain RAG, and how to ship the result as a pinned, evaluated and reversible release.
50 min readWhat a Model-Diffing Agent Finds: Style Shifts Overshadow a Narrow Fine-Tune
A model-diffing agent audited a finance LoRA fine-tune: 82% of runs caught the length and structure shift, while the intended reasoning change stayed hidden.
15 min readLoRA and QLoRA Explained: Fine-Tuning a 7B Model Without 100 GB of VRAM
Why full fine-tuning needs 80 to 100 GB, how low-rank adapters shrink trainable weights to 1.6%, how to pick rank and target layers, and what QLoRA adds.
5 min readEight Linear Algebra Ideas Behind Embeddings, LoRA and PCA
Vectors, matrices, eigenvectors, cosine similarity, norms, SVD and rank explained with small house-price examples and the AI techniques that depend on them.
7 min readHow QLoRA Fine-Tunes Large Models on Small GPUs: 4-bit Basics
Understand QLoRA from the ground up: why full fine-tuning is expensive, how LoRA adapters and 4-bit quantization combine, and when this approach fits your project.
8 min readIs a Bigger LoRA Worth It? Measure Accuracy Against Cost, Not Vibes
A small harness that treats LoRA recipes as data and reports accuracy, training time, latency and memory, so rank and iteration choices rest on numbers.
5 min readTwo LLM workloads on one GPU: vLLM LoRA, agents, and Azure ML fine-tunes
Serve interactive and batch agent traffic from one card with demand LoRA, ClickHouse analytics, Blob artifacts, and Unsloth training tracked in MLflow.
8 min readServing a LoRA Fine-Tune Locally: Verify, Fuse and Avoid Silent Failures
Prove a LoRA adapter actually improved a small model, fuse and serve it behind a local OpenAI-compatible API, and catch the failures that return confident wrong output.
12 min read
About these articles
Request a 24h estimate
Need the same stack in a production operator layer? Send the brief — estimate within 24 hours.