
Clarity is a video captioning agent built for the AMD Developer Hackathon ACT II — Track 2. Given a short video clip, it generates one caption in each of four distinct styles: formal, sarcastic, humorous-tech, and humorous-non-tech. The core design challenge in this track is that asking a model to "watch and be funny" in a single step tends to produce captions that are creative but factually wrong. Clarity solves this with a three-phase architecture that separates seeing from writing: Phase 1 — Perception: The clip is downloaded, and roughly 8 evenly spaced frames are sampled across its duration. These frames are sent to a vision-capable model on Fireworks AI, which produces a purely factual scene description — no jokes, no style, just what's actually visible. Phase 2 — Style Generation: That factual report is passed into four separate, focused calls to a language model on Fireworks AI, one per requested style. Because every style call is grounded in the same facts, the humor and sarcasm never drift into inaccurate territory. Phase 3 — Integration: An orchestrator runs both phases for every task, guarantees every requested style is present even under partial failure, validates the JSON schema, and writes the final results file. A key engineering finding along the way: the available models are heavy reasoning models that produce long internal "thinking" text before answering. Plain prompt instructions weren't reliable enough to control this, so Clarity uses Fireworks' JSON mode to grammar-constrain every response directly into clean, parseable output — a much more robust fix than parsing free text. The agent is fully containerized, builds for linux/amd64 via an automated GitHub Actions pipeline, and was tested against the three official example clips plus additional clips spanning different content types to guard against overfitting.
12 Jul 2026

An AI powered Lawyer app named Legal Mind developed by team Helix AI. Description: Simplify your legal needs with LegalMind. Get instant, accurate advice, review, and drafting right from your phone. Our user-friendly app offers 24/7 support, tailored guidance, and secure, cost-effective solutions for all your legal queries. Use LegalMind today and take control of your legal matters with confidence! LegalMind is an innovative AI-powered legal assistant designed to simplify and enhance your legal experience. Built using Streamlit on Hugging Face and powered by the advanced AI71 API, Whether you need help with contracts, disputes, or everyday legal questions, LegalMind makes navigating the legal landscape straightforward and accessible. Meet our team: Hashir Ehtisham - Leader of team (AI & ML Developer) Huba Iftikhar - Software Engineer Muhammad Tayyab - Programmer
7 Aug 2024