Guides, research notes and what we learn building models.
A total cost of ownership framework for RAG pipelines covering infrastructure, engineering, maintenance, and compliance costs across small, medium, and large deployments.
Read the article
How small amounts of noisy, mislabeled, or low-quality training data disproportionately degrade fine-tuned model performance — and what the research says about the asymmetric impact of bad data.
Read the article
Active learning uses your model to suggest labels, then domain experts confirm or correct. It cuts labeling time by 75% — and when the model runs locally, zero data leaves your infrastructure.
Read the article
Most AI agencies resell API calls and settle for 50-60% gross margins. Agencies running fine-tuned models on their own infrastructure clear 90%.
Read the article
Stop guessing your margins. This calculator breaks down exactly what you're spending per client on API calls vs fine-tuned models — and shows the crossover point where fine-tuning pays for itself.
Read the article
How AI agencies price client-specific agents as monthly retainers: three tiers, the margin math behind each, and what the revenue looks like at scale.
Read the article45-50% of AI SaaS will use hybrid pricing by 2028. Here's how to price AI features — subscription, usage-based, or hybrid — so your margins survive as users scale.
Read the articleYou don't need $10K/month in API costs to ship AI features. Here's the complete stack — fine-tuned model, Ollama, $30 VPS — that runs a production AI SaaS for under $50/month.
Read the articleTraditional SaaS holds 80-90% gross margins. API-dependent AI SaaS averages 25-60%. Five margin traps behind the gap, and how one team reached 82%.
Read the articleBenchmarks for DeepSeek R1 Distill and fine-tuned Llama 3.3 across reasoning, domain accuracy, code and tool calling, plus a routing plan for running both.
Read the articleFive stages that turn enterprise PDFs, email and Slack exports into retrieval-optimized chunks for RAG: ingest, clean, structure, chunk, export.
Read the articleA DPO dataset is prompt, chosen and rejected. Here is where enterprise preference pairs come from, and the full workflow for preparing them on-premise.
Read the articleAugust 2, 2026. That's the deadline for EU AI Act high-risk system compliance. If your AI data pipeline doesn't have audit trails and documentation today, here's the 5-month sprint to get there.
Read the articleA CrewAI crew with four agents can hit 20 or more API calls per task. Fine-tuned local models cut that crew's bill from $5,625 to $581 a month.
Read the articleRouting, tool selection, and response generation are most of a LangGraph agent's LLM calls. Fine-tuned models handle them and cut the API bill 80 to 90 percent.
Read the articleMeasured latency and accuracy for a fine-tuned Gemma 3 4B on iPhone, Pixel, Raspberry Pi and browser, plus QLoRA settings and GGUF quantization picks.
Read the articleFine-tune Phi-4 14B on 12 GB of VRAM with QLoRA. Quantization tables with RTX 4090 tokens per second, plus accuracy against Llama 3.1 and Qwen 2.5.
Read the articleQwen 2.5 covers 29 languages. Fine-tune one model for multilingual classification, support and document extraction, with no separate model per language.
Read the articleJSON mode gets you valid JSON. Fine-tuning gets you guaranteed schema compliance on every field, every type, every time. Here is how to train for it.
Read the articleGPT-4o voice agents run $400 to $1,200 a month at 1,000 calls. See how a fine-tuned local model drops the LLM bill to near zero and answers faster.
Read the articleFine-tuning, RAG, and cloud APIs each solve different problems. Here's a practical routing framework for choosing the right approach per request — and how to combine all three in one system.
Read the articleLoRA adapters run 50-200MB each, so a single base model can hot-swap a personal adapter per user request without multiplying inference costs.
Read the articleThe era of building SaaS on third-party AI APIs is ending. Here's the post-API architecture — fine-tuned local models, GGUF deployment, and zero per-token costs — that makes AI features profitable.
Read the articlePrompt caching cuts costs 60-90% for repetitive context. Fine-tuning eliminates per-token costs entirely. Here's how to know when you've outgrown caching and should fine-tune instead.
Read the article