Model Distillation: Terms of Service and IP Law
What OpenAI, Anthropic, Mistral and xAI terms of service prohibit on model distillation, why copyright doesn't apply, and where open licences leave you free.
Read the articleNotes from
the work.
Guides, research notes and what we learn building models.
Shipping AI Search in Your SaaS Without Per-Query API Costs
A step-by-step tutorial for building natural language search using a fine-tuned 3B-7B model. Includes training data sourcing, model selection, GGUF deployment via Ollama, and latency benchmarks.
Read the article
What Happens When Your AI Provider Cuts You Off? A Survival Guide
Anthropic banned 24,000 accounts overnight. OpenAI deprecated GPT-4o with 2 weeks notice. Your AI provider can change the rules at any time. Here's your survival guide for vendor dependency.
Read the article
Anthropic DeepSeek Distillation: Own Your Models
Anthropic caught DeepSeek, Moonshot, and MiniMax using 24,000 accounts to distill Claude. What that says about building on AI you do not own.
Read the article
From API-Dependent to Model Owner: A 90-Day Migration Playbook
A phased, risk-managed plan for migrating your AI workloads from cloud APIs to fine-tuned models you own. Week-by-week breakdown with concrete milestones for each phase.
Read the article
Build vs. Rent: The True Cost of API-Dependent AI in 2026
The API invoice only tells half the story. When you add deprecation migrations, prompt engineering hours, outage costs, and variable pricing risk, self-hosted fine-tuned models break even in 2-4 months.
Read the articleHow to Distill Open-Source Models Legally
Distillation is legal when the teacher model's licence allows derivative works. How to pick an open-source teacher, train a student, and export to GGUF.
Read the articleFine-Tune a Support Bot for Your Lovable App (No API Costs in Production)
Build an AI support bot that actually knows your product — trained on your docs, your tickets, your tone. Then run it locally for zero ongoing API costs.
Read the articleFine-Tuned AI for Financial Document Analysis: Contracts, Reports, and Filings
How fine-tuned models automate extraction from loan agreements, earnings reports, 10-K filings, insurance policies, and trade confirmations. Includes accuracy benchmarks, cost math, and training data format.
Read the articleFine-Tuned AI for SaaS Customer Support Automation
Your RAG chatbot resolves 34% of support tickets. Fine-tuning pushes that to 87%. Here's how to build a support automation pipeline that actually works — with real numbers on resolution rates, cost per ticket, and the training data you need.
Read the articleFine-Tuned Models for Medical Coding and Clinical Documentation
How to fine-tune local AI models for ICD-10/CPT code suggestion and clinical documentation improvement — covering training data structures, accuracy targets, EHR integration, and ROI math for healthcare organizations.
Read the articleFine-Tuned Tool Calling for n8n and Make.com Workflows
Replace the OpenAI node in your n8n or Make.com workflow with a fine-tuned local model. Same tool routing, same structured output, zero API cost. Here's the exact pattern — from extracting training data from workflow logs to deploying via Ollama.
Read the articleFine-Tuned vs. RAG for Clinical Decision Support: When Each Wins
RAG or fine-tuning for healthcare AI? The answer depends on the clinical task. This guide compares both approaches across 8 healthcare use cases, covering accuracy, latency, cost, HIPAA implications, and a hybrid architecture that combines the best of both.
Read the articleHIPAA-Compliant LLM Fine-Tuning for Healthcare
How to map HIPAA to every stage of an LLM fine-tuning pipeline: de-identification, clinical training data formats, model choice, and on-premise cost math.
Read the articleFine-Tuning vs. Distillation: Which One Actually Gives You an AI Moat?
Distillation copies generic capabilities from larger models. Fine-tuning on your own data creates domain-specific capabilities nobody else has. One gives you a cheaper clone — the other gives you a competitive moat.
Read the articleFunctionGemma and the Rise of Dedicated Tool-Calling Models
Google released FunctionGemma — a 270M parameter model fine-tuned exclusively for function calling. It's tiny, fast, and signals a major shift: the era of task-specific models is here. What this means for builders, when to use it, and when to fine-tune your own.
Read the articleLoRA Adapters Per Healthcare Specialty: Radiology, Pathology, Primary Care
How to serve multiple hospital departments from a single base model using specialty-specific LoRA adapters. Covers architecture, training data requirements, storage math, adapter management, and performance benchmarks.
Read the articleIs Model Distillation Legal? The Three Levels
Distillation is standard practice at every major lab. Here is the line between a permitted open-source build, a terms of service breach, and theft.
Read the articleMulti-Step AI Agents on Local Models: Architecture and Patterns
Multi-step reasoning is where small models historically fail. But with the right architecture — specialist chains, scratchpad fine-tuning, and hybrid routing — you can run reliable multi-step agents on 7B-14B models locally. Here are three proven patterns with benchmarks.
Read the articleMulti-Tenant Fine-Tuning: Per-Customer AI Models in Your SaaS
Your SaaS customers want AI that understands their data, not generic responses. Here's how to architect per-tenant fine-tuned models using LoRA adapters — with real storage math, cost breakdowns, and a serving architecture that scales to hundreds of tenants.
Read the articleOpenAI Deprecated 5 Models in 6 Months — Here's What It Cost Businesses
GPT-4o, DALL-E-3, Assistants API, Realtime API — OpenAI deprecated them all within months. Each deprecation forces engineering migrations that cost far more than the API bill itself.
Read the articlePer-Client AI Agents for Agencies: LoRA + Tool Calling Playbook
Every agency client gets the same GPT wrapper. That's the problem. With per-client LoRA adapters and custom tool schemas, you can deliver AI agents that know each client's CRM, workflows, and terminology — at 50-200MB per client. Here's the full playbook.
Read the articleHow to Price AI Features in Your SaaS: Usage-Based vs. Tier-Included
Four pricing models for AI features in SaaS, the margin math behind each, and why most teams pick the wrong one. Includes real examples from Notion, Linear, and Intercom.
Read the articleBuilding Reliable AI Agents with Fine-Tuned Local Models: Complete Guide
Most AI agents are just GPT-4 wrappers — expensive, unreliable at scale, and dependent on cloud APIs. Fine-tuned local models hit 98%+ accuracy on your specific tools at zero per-query cost. Here's the complete architecture.
Read the article