CyberFreezeDev
Loading
Blog series
Practical guides to running models on your own hardware, building retrieval and agents, learning the basics, and choosing and securing AI for a business. Every guide cites the documentation, papers and official sources it relies on.
This series continues the Cyber Freeze AI blog that ran at ai.cyberfreezedev.com. Its articles have been rewritten here with current tools, versions and sources, and the old addresses now lead to the new guides.
Install Ollama and run deepseek-r1:8b, a 5.2 GB distilled model that fits an 8 GB GPU. Here is which size suits which hardware, how to install on each system and when local is the wrong call.
A 4-bit model of up to about 8 billion parameters runs well on a CPU, because speed depends on memory bandwidth, not cores. Here is how much RAM each size needs and what speed to expect.
Memory decides which model a GPU can run and bandwidth decides how fast. Here is what 16, 24, 32, 48 and 96 GB cards hold, how consumer, workstation and data centre cards differ, and what to check when buying in India.
GPU memory for an LLM is the weights plus the KV cache plus a runtime allowance, and all three can be worked out from the model's config file. Here is the arithmetic for 8B, 32B and 70B models and what fits on 8, 16, 24, 48 and 80 GB cards.
GGUF is a file format for llama.cpp and Ollama, while AWQ and GPTQ are 4-bit methods that GPU servers such as vLLM load. Here is what each one is, what names like Q4_K_M mean and how to choose.
Ollama suits a laptop, llama.cpp suits one machine you want full control of, and vLLM suits a GPU server that many people use at once. Here is how they differ on batching, model formats, APIs and licences.
Open image models can run on a single GPU in your office, but several popular ones are not free for business use. Here is what the licences say, what hardware you need and when a hosted tool is simpler.
Seven types of retrieval augmented generation, where each name comes from, and a table of which one fits FAQs, manuals, contracts, archives and questions that span several systems.
Seven places retrieval augmented generation earns its keep in a business, with what goes in, what comes out, and the one thing that usually breaks in each.
Five ways to split documents for a RAG system, what LangChain, LlamaIndex and published evaluations say about chunk size and overlap, and which to start with for policies, contracts, manuals and transcripts.
Why there is no single best embedding model for RAG, what to read on a model card, how Hindi and Tamil change the choice, and a short test you can run on your own documents.
Vector search finds meaning but misses exact strings like part numbers and error codes. Here is how BM25 keyword search, reciprocal rank fusion and a reranker fix that, and which databases support it out of the box.
Where a reranker sits in a RAG pipeline, when it fixes retrieval and when it cannot, six open reranker models with licences checked on their model cards, and what drives the extra latency.
A RAG chatbot is only as trustworthy as the test you run it against. How to build a test set from real tickets, which retrieval and answer metrics to track, and how to stop a change from quietly breaking it.
You do not need a degree or a paid bootcamp to learn AI, but you do need an order to learn it in. Six stages, the free official course for each, where different roles can stop, and how a team can learn together.
Machine learning uses four areas of maths, but you need far less of each than a university syllabus suggests. What each one does in practice, how much is enough for your role, and the free books and courses worth your time.
Most of the work in a machine learning project is getting business data into shape. How to load, clean, join and check CSV and Excel exports with pandas 3.1, with a short example you can adapt.
Linear regression is the simplest forecasting model worth knowing. A worked sales example in scikit-learn, how to read what it prints, and the situations where it gives confident wrong answers.
Most small businesses should try a spreadsheet or a written rule before machine learning. Here is what forecasting, classification and anomaly detection do, the data each needs, and the signs you are not ready yet.
An n8n AI agent is the AI Agent node with a chat model, tools and optional memory. Here is a worked example that triages support email into a CRM, with a person approving every write.
An AI agent should propose and a person should approve wherever an action moves money, reaches a customer or deletes data. Here is how to build the queue, what to do when nobody answers, and what to log.
You can run several AI agents on your own hardware with open-weight models, and none of your documents has to leave the building. Here is the architecture, which frameworks and servers work fully local, what to switch off, and why small local models drift on long plans and how to stop it.
No model is best for every business. Here is how to shortlist GPT, Gemini, Claude or an open-weight model on data, cost, context, tools and languages, using each vendor's own documentation.
Sixteen questions to put to an AI developer or agency before you hire, grouped by LLMs and RAG, data and evaluation, deployment and cost, and security, with what a good answer contains.
OWASP published the 2026 edition of its Top 10 for LLM Applications on 3 August 2026. Here is each risk in two sentences, with a chatbot example and the control that addresses it.
No filter stops every prompt injection, because a language model cannot tell instructions from data. Here is what OWASP, the NCSC and NIST say, and the design rules that keep a fooled chatbot from doing harm.
A convincing deepfake can beat your eyes and ears, so the defence is a process: warning signs, a call back on a number you already hold, and payment rules no single call can override.
