PointBlank

A developer built PointBlank, an open-weight-model-powered AI speech coach that enforces the PREP framework (Point, Reason, Example, Point) under a 20-second prep and 90-second speaking limit, grading users on cadence, filler words and structure and supplying a rewritten version of their answer. The React/Vite and Tailwind CSS v4 app, built for a friend who rambles in technical interviews, uses the Web Speech API and a localStorage-based daily streak counter. The developer argues open-weight models like Gemma are essential because closed APIs' alignment layers flag harsh coaching personas as safety violations.

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend https://dev.to/challenges/hacktoberfest-weekend-2026-10-01 I built PointBlank, an unforgiving, brutally honest AI speech coach designed in a strict Neo-Brutalist aesthetic. I built this for a close friend who is an incredibly talented developer but struggles with "rambling" during high-stakes technical interviews and executive design reviews. Whenever they get nervous, they stack qualifiers, use filler words "um," "like" , and take forever to get to the core point. PointBlank solves this by enforcing the PREP Framework Point, Reason, Example, Point under a strict time constraint. It gives you exactly 20 seconds to mentally organize, and a hard 90 seconds to speak. It records the speech and uses an AI model to ruthlessly audit the cadence and structure. If you ramble, the AI issues a failing grade, calls out filler words, and provides a polished rewrite of what you should have said. Test your skills on PointBlank: Turn on your mic, start the clock, and let the AI roast your impromptu speaking skills. It won't hold back. PointBlank solves this by enforcing the PREP Framework Point, Reason, Example, Point . It records their speech and uses an AI model to ruthlessly audit their cadence and structure. If they ramble, the AI issues a failing grade, calls out their filler words, and provides a polished rewrite of what they should have said. Live App: https://point-blank-alpha.vercel.app/ https://point-blank-alpha.vercel.app/ GitHub Repo: https://github.com/jayantkumar27/PointBlank https://github.com/jayantkumar27/PointBlank PointBlank is built using React Vite and Tailwind CSS v4 to achieve a high-contrast, zero-border-radius Neo-Brutalist UI that… PointBlank is built using React Vite and Tailwind CSS v4 to achieve a high-contrast, zero-border-radius Neo-Brutalist UI that matches the harsh nature of the coach. I also wired up a zero-login daily streak counter using localStorage to keep the pressure on. For the AI pipeline: Open innovation and open-weight models like Gemma are critical for a project like PointBlank. Closed APIs often have heavy alignment layers that mistakenly flag "brutally honest" or "harsh" personas as violations of safety policies, watering down the feedback into polite, unhelpful suggestions. By building around open-weight models, developers have the freedom to construct specific, highly-tuned agent personas that serve a real behavioral conditioning purpose without being restricted by a corporate safety filter. This project was built iteratively with the assistance of an AI coding agent that helped translate high-fidelity neo-brutalist design mockups into raw Tailwind components, wired up the Web Speech API, and structured the GenAI evaluation prompts.