Skip to content
View amirfandev's full-sized avatar

Achievements

Achievement: Pair ExtraordinaireAchievement: Quickdraw

Achievements

Achievement: Pair ExtraordinaireAchievement: Quickdraw

Block or report amirfandev

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
amirfandev/README.md

Hi, I'm Irfan. I build reliability tooling for LLM systems: schema gates that refuse bad output, failure taxonomies that give bugs names, eval harnesses whose numbers you can recompute yourself. By day that means agent pipelines that turn requirements into reviewed, tested changes, with validation gates between every stage. The open-source work is the same discipline with all the evidence published.

How I build

  • Runnable over described. If you can't clone it and run it, it's a claim, not a tool.
  • Numbers carry their evidence. An eval result without the raw data behind it is an anecdote with a decimal point.
  • Failures get names. "Wrong" is not a category. "Unit confusion in numeric fields" is. A named failure class with a test attached can be counted, fixed, and watched for regressions.
  • One problem per repo. Small tools compose. Platforms rot.

None of this is a new habit. Thirteen years of iOS and real-time media came first, including a 60fps camera app that passed 8 million downloads and the original Flashlight app at number one in App Store Utilities. Then four years as tech lead on a healthcare EMR platform in Abu Dhabi, where the systems carried e-prescriptions and I ran the internal HIPAA and ADHICS audits. Hard gates stopped being a style preference somewhere in there.

If something breaks, or a number doesn't recompute, open an issue. That is the fastest way to reach me. And if you are hitting agent failure modes that don't have names yet, I want to hear about those too: LinkedIn.

Pinned Loading

  1. compaction-guard compaction-guard Public

    Catches safety constraints and task invariants that an LLM agent's context compaction silently drops, then repairs them. Zero dependencies.

    Python

  2. toolwarden toolwarden Public

    Deterministic policy gate for LLM agent tool calls. Checks every call against policy before it executes, allow or deny by code not by a model, with an audit trail. Zero dependencies.

    Python

  3. load-board-matcher load-board-matcher Public

    Extracts a driver profile from a call transcript behind a validating schema gate, then screens a load board before ranking by effective rate per mile.

    Python