On-Device Text Classification for Mobile Apps
How to build fast, accurate text classification that runs on the user's phone. Sentiment analysis, content categorization, intent detection, and spam filtering without an API call.
Read the articleNotes from
the work.
Guides, research notes and what we learn building models.
OpenAI API for Mobile Apps: Quick Start and the Costs Nobody Mentions
A practical guide to integrating OpenAI's API into iOS and Android apps, with honest cost projections at 1K to 100K users that most tutorials skip.
Read the article
What Happens When OpenAI Deprecates the Model Your App Depends On
Model deprecation is not hypothetical. OpenAI has deprecated 15+ models since 2023. When your app depends on a specific model version, deprecation means a forced migration under a deadline you did not choose.
Read the article
OTA Model Updates: Keeping Your On-Device AI Current
How to push model updates to users without an app store release. Version checking, background downloads, rollback strategies, and the infrastructure for over-the-air model delivery.
Read the article
Phi-4 Mini for Mobile: Microsoft's Small Model on iOS and Android
Microsoft's Phi-4 Mini packs strong reasoning into 3.8B parameters with an MIT license. How it compares to Llama and Gemma for mobile deployment, and when to choose it.
Read the article
Q4 vs Q5 vs Q8: GGUF Quantization on Mobile
How Q4, Q5, and Q8 GGUF quantization change model size, speed, quality, and memory use on iPhone and Android, measured on real devices.
Read the articleShipping GGUF Models: App Store Bundling vs Post-Install Download
Two ways to get your GGUF model onto the user's device. Bundle it with the app for simplicity, or download post-install for flexibility. Architecture, size limits, and best practices for both.
Read the articleBuilding a Training Dataset from Your App's User Interactions
Your app already generates the training data you need for fine-tuning. How to collect, clean, and format user interactions into a dataset that produces a high-quality on-device model.
Read the articleYour User's Data Leaves Their Phone on Every AI Request
Every cloud AI API call sends user data to a third-party server. What that means for privacy, compliance, user trust, and your app's long-term viability.
Read the articleFine-Tuned 3B vs GPT-4: Why Smaller Models Win at Domain Tasks
Academic research shows fine-tuned 3B-7B models consistently beat GPT-4 on domain-specific tasks. Here's the evidence, the pattern, and how to apply it in your app.
Read the articleFine-Tuning for App Developers: A Non-ML-Engineer's Guide
A practical guide to fine-tuning AI models for mobile app developers. Learn LoRA, QLoRA, and GGUF export without needing an ML background.
Read the articleRun GGUF Models on iOS and Android with llama.cpp
Package a fine-tuned model as GGUF and run it with llama.cpp on iOS and Android: quantization choices, real file sizes, throughput, and delivery patterns.
Read the articleHow Many Training Examples Do You Actually Need? The 100-Sample Myth
The real data requirements for fine-tuning AI models. Research shows 50-500 examples can be enough for many tasks. Here's what the papers say and how to build your dataset.
Read the articleOn-Device vs Cloud API: The Real Math at 10K, 50K, and 100K MAU
A no-fluff cost breakdown of cloud API pricing vs on-device inference at scale. See exactly when on-device fine-tuning pays for itself, with tables, real pricing data, and the hidden costs nobody puts in the README.
Read the articleThe AI Data Quality Framework: Measuring What Actually Matters for Training Data
A systematic framework for measuring and ensuring AI training data quality across five dimensions, with scoring methodology and maturity levels for enterprise teams.
Read the articleAnnotation Quality Metrics Beyond Cohen's Kappa: A Practitioner's Guide
A rigorous guide to annotation quality metrics beyond Cohen's Kappa — covering Krippendorff's Alpha, F1 agreement, confusion matrices, calibration sessions, and when to use each.
Read the articleData Preparation Time Estimator: How Long Does AI Data Prep Take by Document Type
A time estimation framework for AI data preparation by document type and volume. Compare manual vs automated processing times for PDFs, Word docs, Excel files, scanned documents, and more.
Read the articleThe Data Quality Maturity Model for Enterprise AI: Where Does Your Team Stand?
A 5-level maturity model for enterprise AI data quality — from Ad-hoc to Optimized — with assessment criteria, metrics, and tooling recommendations at each level.
Read the articleHow to Define Data Quality SLAs for AI/ML Service Engagements
A practical guide and template for AI/ML service providers to define data quality SLAs with clients — covering what to promise, how to measure, what to exclude, and remediation terms.
Read the articleEU AI Act Compliance Readiness Checker for Data Pipelines
A compliance readiness framework for EU AI Act Articles 10, 12 and 19 applied to AI training data pipelines. Includes checklist tables for high-risk and limited-risk systems with the August 2026 deadline in focus.
Read the articleThe Five Dimensions of AI-Ready Data Quality: A Scoring Guide
A detailed scoring rubric for evaluating AI training data across five dimensions — Completeness, Consistency, Accuracy, Timeliness, and Relevance — with concrete enterprise examples at each level.
Read the articleNode-Graph Pipeline vs Python Scripts for RAG: When Visual Wins and When It Doesn't
Visual pipeline builders and Python scripts are both valid ways to build RAG. But they optimize for different things — and choosing wrong costs you maintenance burden or flexibility. Here is when each approach fits.
Read the articlePII Exposure Risk Scorecard: Self-Assessment for AI Pipelines
A self-assessment scorecard with 10 scored risk factors for evaluating PII and PHI exposure in your AI data pipelines. Score your risk level and identify gaps before they become incidents.
Read the articlePII in Vector Stores: Why Embedding Sensitive Data Is a Compliance Liability You Cannot Undo
Once personal data is embedded as a vector, it cannot be selectively deleted, redacted, or audited. Every query against that vector store potentially surfaces PII. The only safe approach is to redact before you embed.
Read the article