Guides, research notes and what we learn building models.
Total cost of ownership analysis for AI features from seed to Series B. Real math on the cost cliff, hidden multipliers, break-even points, and why investors care about AI margin.
Read the article
Your AI features work. Your API bill is growing faster than revenue. Here's the decision framework, cost math, and migration path for moving from per-token APIs to fine-tuned models — with real numbers at every step.
Read the article
A practical guide to generating synthetic training data for fine-tuning — covering prompt strategies, quality filtering, distribution matching, and the 80/20 rule for mixing real and synthetic data.
Read the article
The biggest gap in fine-tuning guides: nobody covers how to actually build the dataset. Here's a step-by-step process to create tool-calling training data — from schema documentation to synthetic expansion to JSONL formatting — with real examples for a 5-tool customer service agent.
Read the article
Turn a client's support tickets, chat logs, sales calls and failure logs into an evaluation dataset, then label, size and maintain the eval set over time.
Read the article
Export your n8n execution history, fine-tune a model on it, run it on Ollama, and swap out the OpenAI node. Your per-execution API bill goes to zero.
Read the articleUnderstanding and fixing hallucination in fine-tuned models — why fine-tuning can make hallucination worse, detection techniques, and practical mitigation strategies for production deployments.
Read the articleYour Lovable prototype works and users are signing up. Five gaps stand between it and production: AI cost, reliability, privacy, latency, and lock-in.
Read the articleHow to safely migrate from cloud AI APIs to a fine-tuned model by running a production A/B test. Covers routing architecture, what to measure, statistical significance, and the gradual migration path from 10% to 100%.
Read the articleA decision framework for AI agencies choosing between cloud GPU rentals, consumer hardware purchases, and dedicated inference chips. Includes break-even analysis, client volume thresholds, and compliance considerations.
Read the articleFinancial services firms want AI but can't use cloud APIs due to SOC 2, PCI-DSS, and FINRA constraints. This is a massive opportunity for agencies that can fine-tune and deploy models on-premise. Here's how to capture it.
Read the articleAI agents that depend on cloud APIs are fragile, expensive, and privacy-risky. Fine-tuned tool-calling models running on edge hardware create agents that work offline, respond instantly, and keep data local.
Read the articleCPUs went from ENIAC to smartphones in 60 years. AI inference is following the same arc — from cloud data centers to dedicated silicon to on-device chips. Fine-tuning is the software layer that makes each hardware generation useful.
Read the articleCost per 1M tokens for cloud APIs, self-hosted GPUs and dedicated silicon like Taalas HC1, priced out across agency, indie app and hospital workloads.
Read the articleSOC 2, PCI-DSS and FINRA rules turn cloud AI APIs into a compliance risk for banks. Here is why fine-tuned models running on-premise pass review.
Read the articleThe edge AI hardware market is projected to hit $59 billion by 2030 and 80% of inference is expected to happen locally. Here's what's driving the shift, what hardware is emerging, and why fine-tuning is the missing piece.
Read the articleA comprehensive guide to deploying fine-tuned AI models in financial services. Covers SOC 2, PCI-DSS, and FINRA compliance, five production use cases, and why on-premise fine-tuned models are replacing cloud APIs in banking and finance.
Read the articleWhat each M-series Mac can run, and how to deploy a fine-tuned model locally with Ollama, MLX or llama.cpp at zero cost per query.
Read the articleA 10-point quality checklist for agencies and teams deploying fine-tuned models to clients — covering accuracy benchmarks, hallucination detection, format compliance, latency, and safety guardrails.
Read the articleGeneric models hallucinate function names and mangle parameters. Fine-tune a small model on your own tool schema for reliable tool calls at near-zero cost.
Read the articleTaalas, Apple and Qualcomm are all adding LoRA adapter support to their AI silicon. Here is the engineering and business logic behind the convergence.
Read the articleYou're paying frontier-model prices for what amounts to pattern matching and JSON generation. A fine-tuned 8B model handles tool calling at 90%+ accuracy for zero per-query cost. Here's the math and the migration path.
Read the articleHow many MB a LoRA adapter takes at r=4 through r=128 on an 8B model, the formula to estimate it before you train, and the right rank per device.
Read the articleHow Taalas's HC1 chip, Tether Data's QVAC Fabric LLM and LoRA-Edge research turn LoRA adapters into the deployment interface for silicon and edge devices.
Read the article