Simple Request Router

This project delivers an intelligent routing agent engineered to balance strict enterprise accuracy requirements against cloud API token consumption. Built to comply with an evaluation environment restricted to a 4 GB RAM and 2 vCPU footprint, the system relies on an intentional two-step isolation architecture that separates intent classification from task execution. At the core of the local optimization strategy is an ultra-lightweight, 4-bit quantized Gemma 4 (Effective 2B) parameter model running locally via llama-cpp-python. When an evaluation payload enters the pipeline, the local model acts strictly as a zero-token-cost binary classification gate. It evaluates incoming prompts across eight distinct capability domains. Direct, low-complexity natural language tasks - such as sentiment analysis, named entity recognition, basic definitions, and text summarization - are processed and answered entirely in-house, incurring a total external token cost of zero. Conversely, when the local gate detects complex algorithmic patterns, multi-step mathematical reasoning, code debugging specs, or deductive logic puzzles, it flags the task as COMPLEX and routes it to the premium Fireworks AI cloud infrastructure. By isolating the routing logic and utilizing defensive system constraints like strict stop-token structures and max-token caps, this architecture minimizes proxy token overhead while ensuring maximum accuracy for more demanding requests.

SovereignSwarm: Auto Multi-Tone Video Captioner

SovereignSwarm is an advanced, high-throughput video intelligence application designed to automate the process of localized multi-tone video captioning. Built purely on the fast Fireworks AI inference ecosystem, the system acts as a seamless bridge between computer vision and generative linguistics. The pipeline utilizes OpenCV to dynamically extract key chronological frames from any short video under 30 seconds. These frames are processed through a cutting-edge Vision-Language Model to generate a raw, highly factual description of the scene. The architecture then leverages an asynchronous text-generation layer to rewrite the core description into four completely distinct narrative styles: Formal, Sarcastic, Humorous Tech, and Everyday Humor. The final payload is formatted and returned as a strictly validated, production-ready JSON schema. This ensures seamless integration into enterprise content management workflows, video editing suites, or automated social media pipelines without requiring heavy local infrastructure or complex cloud configuration.

Lexyprep

Verity — Fact-Checking for the Rest of Us

Most people believe first and verify never , there simply isn't time to check every headline, post, or forwarded message before deciding whether to trust it. That gap is where misinformation lives and spreads. Verity closes it. You give Verity a claim — a headline, a statistic, a line from an article — and it separates what's actually checkable from what's just opinion, then searches for independent evidence across 40+ public authoritative sources. Instead of trusting a single model's best guess, Verity requires independent checks to agree before returning a verdict: true, false, or somewhere in between, with a source trail you can click through and verify yourself. Every verification runs live on AMD Instinct hardware via the Fireworks AI API, using Gemma and other open-source models routed through a 5-pass consensus architecture — nothing pre-computed, nothing faked. The models are swappable and inspectable by design, because a trust tool that asks you to trust it blindly defeats its own purpose. Verity is built for the moment right before someone shares something they haven't checked , turning fact-checking from a newsroom luxury into something anyone can do in seconds.

Aprendia App

Getting an answer today takes seconds but answering fast is not the same as learning. Students read an answer and a week later it's gone. Aprendia is a platform that turns any topic into a complete, playable course built on real learning science, so knowledge actually sticks. Every topic becomes a campaign: chapters and missions with a final challenge you only clear by mastering the content. Aprendia has 13 interactive exercise types ordering, matching, image hotspots, diagram completion, video that pauses and quizzes you, even open answers graded by an AI judge with feedback. Reviews are scheduled with FSRS spaced repetition, give it your exam date and a Readiness Index predicts how prepared you'll be, with full mock exams. A RAG tutor answers using only your course material and shows citations if something isn't in the sources, it says so. Courses can anchor to your own material: PDFs, links, YouTube videos, audio transcribed locally, or a cited web-research pass. Gamification is ethical (XP by mastery, never screen time) and an assisted mode makes every screen narrated and fully keyboard-driven. All of Aprendia's AI runs on AMD hardware through one provider-agnostic layer: Fireworks AI serves open models (GLM, gpt-oss) on AMD GPUs to generate full courses in 20 minutes for ~$0.25, while our self-hosted Gemma 3 27B vLLM + ROCm on an AMD Instinct MI300X in AMD Developer Cloud powers the live judge and tutor with no rate limits. The full code is open (MIT) on GitHub, and one command seeds a complete demo: 7 campaigns, demo users, real progress no AI key required.

Video Captioner

Video Captioner V2 is a 5-stage AI pipeline that generates accurate, multi-style captions for silent videos by analyzing visual content with Gemma 4 31B vision model. The Problem: Most captioning tools rely on audio transcription, which fails for silent videos. Track 2 requires describing what's IN the video visually. Our Solution: - Stage 1: Adaptive Frame Extraction (9-15 frames based on duration) - Stage 2: Vision Analysis with Gemma 4 31B - Stage 3: Frame-to-frame comparison for scene changes - Stage 4: Thinking Board documentation - Stage 5: Caption generation + scoring + re-generation for quality Caption Styles: Formal, Sarcastic, Humorous-Tech, Humorous-NonTech Results: 6/6 videos processed, 97s average, auto-scoring ensures quality.