Projects

Things I've built to solve real problems — for myself, and for the people I work with. After 20+ years building and leading engineering teams at Samsung, Intuit, and Atlassian, I made a deliberate choice: stay hands-on. I still write code, ship products, and experiment with the same AI tooling I recommend to the nonprofits and teams I advise. These are a few of the things I've built recently — spanning personal knowledge systems, decision-making frameworks, and public tools — all using agentic AI workflows (Claude Code, MCP) as a core part of how I build today.

Lore — AI-Powered Second Brain

An open-source second brain that learns, quizzes, and reflects with you — powered by Claude, owned entirely by you.

Lore is a personal knowledge system built on Claude Code and MCP that turns scattered notes, goals, and reading into a structured, queryable "second brain." It runs entirely in plain Markdown files on your own machine — nothing sent to a server — and uses five agentic commands (/learn, /ingest, /ask, /assess, /reflect) to research topics aligned to your goals, build a linked wiki, quiz you on what you've learned, and write daily journal entries that consolidate decisions and open loops.

Link: github.com/pradnk/lore

Temporal Pincer — A Decision Framework for Complex Projects

A reusable method — and a Claude Code skill — for pressure-testing any outcome-based project before you commit to it.

Temporal Pincer is a multi-agent structured planning method that attacks a problem from two directions at once: a forward pass (what can I actually do next, from today's reality) and a backward pass (assume the outcome already happened — write the retrospective, in past tense, and work out what had to be true). The two passes are written independently, then diffed against each other — the disagreements are usually the things that would have quietly sunk the project. Built out as a Claude Code skill, it runs the entire loop using two isolated AI subagents that never see each other's output, so the forward and backward passes stay genuinely independent, then facilitates the reconciliation with you and hands you a single action-first plan with kill criteria.

Link: github.com/pradnk/ai-temporal-pincer

Sentinel — Intelligent PDF Validator with a Full Eval Harness

A small utility with an outsized point to make: don't ship an AI check without proving it works.

Sentinel is a REST API and web app that validates uploaded PDFs against configurable rules — page count, image count, and language — auto-sorting them into accepted and rejected folders with a plain-English reason. The validation logic itself is simple by design; the real substance is the eval harness built around it: a golden-dataset regression runner with labelled test cases, a blind LLM-judge that never sees the checker's verdict (so disagreements reflect real mistakes, not anchoring), and a GEval-style judge that reasons step-by-step, scores confidence numerically, and re-samples automatically on borderline documents.

Link: github.com/pradnk/sentinel

ResumeForge — AI Resume-to-Job-Description Matcher

An honest, harsh-but-fair resume matcher and ATS-safe resume generator — built two ways, so anyone can use it.

ResumeForge scores how well a resume matches a job description, researches the hiring company, and generates an ATS-safe resume in five formats with live previews and PDF export. It ships as two interchangeable experiences sharing identical scoring logic: a local Next.js web app for anyone with an API key (with score gauges and interactive suggestion review), and a guided conversational skill that runs inside Claude, ChatGPT, or Gemini for anyone who just has a subscription — no API key required.

Link: github.com/pradnk/resume-forge