How much time will I actually save with Fazm?

Most users report saving 45-90 minutes per day on repetitive tasks. The biggest time savings come from email management (drafting replies, sorting, archiving), form filling (expense reports, CRM updates, applications), research (comparing products, gathering data from multiple sites), and social media (cross-posting, scheduling, replying). For example, booking a flight that normally takes 10 minutes of clicking through airline sites takes a single voice command with Fazm. Filing an expense report that requires opening a PDF, extracting numbers, and entering them into a spreadsheet becomes one spoken sentence. The more repetitive and multi-step your daily work is, the more time you get back.

What exactly can Fazm control?

Fazm controls everything you can do manually on your Mac. That includes mouse movements and clicks, keyboard typing, browser navigation and form filling, app switching, file management in Finder, and system-level actions. It uses direct browser DOM control for web interactions - meaning it reads and manipulates the actual page structure rather than taking screenshots and guessing where to click. For native macOS apps, Fazm uses accessibility APIs to interact with buttons, menus, text fields, and other UI elements. This works across any application: Chrome, Safari, VS Code, Slack, Figma, Terminal, Mail, Calendar, and more.

How is Fazm different from ChatGPT or Claude?

Chatbots like ChatGPT and Claude give you answers in a conversation window. Fazm takes action directly on your computer. When you say 'reply to Sarah's email,' Fazm opens your email app, finds Sarah's message, writes a contextual reply, and sends it. A chatbot would tell you what to write, but you still have to do all the clicking and typing yourself. Fazm moves your mouse, types on your keyboard, clicks buttons, and navigates between apps. It is a computer agent, not a chat assistant.

How does the memory layer work?

Fazm builds a personal knowledge graph by extracting information from your files, browsing history, conversations, and daily activity. Over time, it learns your contacts (names, emails, relationships), accounts and credentials, preferences (formatting, tone, scheduling habits), and frequently used workflows. In your first week, you might say 'Reply to Sarah's email, she's my cofounder, her email is sarah@acme.com.' By week four, you just say 'Reply to Sarah' because Fazm already knows who she is. By week eight, Fazm can draft the reply before you even ask. All of this data stays locally on your Mac - your knowledge graph never leaves your machine or gets sent to any cloud service.

Is Fazm safe? Can it mess up my computer?

Safety is built into every layer of Fazm. You watch every action in real-time on your screen - Fazm moves your mouse and types visibly, so there is nothing hidden. You can stop any action at any point by pressing a keyboard shortcut. Fazm also shows you what it plans to do before executing destructive actions like deleting files or sending emails, giving you a chance to confirm or cancel. Because Fazm is open source, the code is fully auditable on GitHub. There is no hidden behavior, no background processes making changes you cannot see, and no actions that bypass your screen.

What platforms does Fazm support?

Fazm currently supports macOS, running natively on both Apple Silicon (M1, M2, M3, M4) and Intel Macs. It is built with Swift and uses macOS-specific frameworks like ScreenCaptureKit for screen analysis and Accessibility APIs for app control, which is why macOS is the primary platform. Windows and Linux support are on the roadmap but do not have release dates yet. You can also control your Mac remotely from your phone using Fazm Remote, which gives you voice access to your desktop while you are away from your computer.

How does Fazm voice input work?

One keyboard shortcut activates push-to-talk. Hold the key, speak naturally about what you need done, and release. Fazm transcribes your speech locally using on-device models, so there is no network latency for the voice recognition step. You do not need wake words like 'Hey Siri' or 'OK Google' - just press and talk. You can also type commands into the floating toolbar if you prefer text input. Fazm understands natural language, so you do not need to memorize specific commands. Say 'book me a flight to Tokyo next Thursday' or 'find the cheapest AirPods on Amazon' and Fazm figures out the steps.

Does Fazm work in languages other than English?

Yes. Fazm works in any language - English, Russian, Spanish, German, Japanese, and more. You can speak or type in your preferred language and Fazm will understand and act. Some of our most active users run their entire workday in Russian, issuing voice commands to manage Jira tickets, write Slack messages, and navigate internal tools. There is no language setting to configure - just speak or type naturally and Fazm handles it.

Is my data sent to the cloud?

Fazm processes screen analysis and memory entirely on your machine. The personal knowledge graph, file indexing, and screen capture all run locally. The only data that leaves your Mac is the intent - a text description of what you want to accomplish, which is sent to the AI model (Claude or GPT) for action planning. Your actual screen content, file contents, browsing history, and personal knowledge graph never leave your computer. Fazm is also open source, so you can verify exactly what data is transmitted by reading the code on GitHub. For users who need complete air-gapped operation, local model support is planned.

Validating LLM Behavior Before Production - Golden Datasets and Automated Evals

Fazm Team·March 18, 2026·2 min read

llm evaluation testing production ai-agents

Stop Pushing LLM Changes Without Validation

Every prompt change is a potential breaking change. Every model upgrade can shift behavior in unexpected ways. If you are pushing these changes to production without automated validation, you are gambling with your users' experience.

What Is a Golden Dataset?

A golden dataset is a curated set of input-output pairs that represent your agent's expected behavior. Each entry says: "given this input, the agent should produce something like this output."

Building one is straightforward:

Start with real interactions. Pull successful agent runs from your logs. These are already validated by user satisfaction.
Cover edge cases deliberately. Add examples for the tricky inputs - ambiguous requests, multi-step workflows, error recovery scenarios.
Keep it small but representative. 50 to 100 well-chosen examples beat 10,000 random ones. Quality of coverage matters more than quantity.

Automated Eval Pipeline

Once you have golden data, the eval pipeline is simple:

Run your agent against every input in the dataset
Compare outputs against expected results using a scoring function
Flag any score below your threshold
Block the deploy if too many examples regress

The scoring function does not need to be perfect. Even a basic semantic similarity check catches most regressions. For structured outputs, exact field matching works well.

When to Run Evals

Before every prompt change - even "minor" wording tweaks
Before every model upgrade - Claude 3.5 to Claude 4 behavior shifts are real
On a schedule - API behavior can drift even without your changes

The Cost of Skipping This

Without evals, you discover regressions through user complaints. By then, the damage is done - broken workflows, lost trust, support tickets. A 5-minute eval run is cheaper than a single angry user.

The teams shipping reliable AI agents all have one thing in common: they treat prompt changes like code changes, with tests that must pass before deploy.

Fazm is an open source macOS AI agent. Open source on GitHub.

Validating LLM Behavior Before Production - Golden Datasets and Automated Evals

Stop Pushing LLM Changes Without Validation

What Is a Golden Dataset?

Automated Eval Pipeline

When to Run Evals

The Cost of Skipping This

More on This Topic

Related Posts

What Breaks When You Evaluate an AI Agent in Production

Affordable AI Agent Evaluation - Recording and Replaying Tool Call Traces

AI Agents Break One Step After the Demo Ends