Three Hot GitHub Projects: AI Video Editing, Local LLM Cluster, and Investment‑Agent

video-use: AI‑driven video editing agent

video-use is an AI video‑editing agent that integrates with Claude Code, Codex and similar tools. Users place raw footage in a folder and issue a natural‑language command such as “make a publish‑ready cut”. The system automatically performs trimming, color grading, subtitle burning, optional animation overlays, and self‑evaluation.

Key capabilities

Automatic filler and pause removal : detects “umm”, “uh” and silent gaps from the audio transcript and cuts them at word boundaries.

Automatic color grading : applies two preset looks (warm cinematic, neutral punch) via ffmpeg chains; custom chains can be supplied.

Automatic subtitles : burns subtitles into the video in configurable 2‑word uppercase chunks; position, font and color are adjustable.

Animation overlay : optional animation layers can be generated by integrating HyperFrames, Remotion, Manim or PIL, each handled by a sub‑agent.

Self‑evaluation : after each editing node the AI runs timeline_view to check for visual jumps, audio spikes or subtitle errors, retrying up to three times.

Design principle

The AI never watches the video; it reads a compact transcript.

Layer 1 – Audio transcription : each clip is processed with ElevenLabs Scribe, producing word‑level timestamps, speaker diarization and event tags. The data are packed into a ~12 KB takes_packed.md file that serves as the AI’s script.

Layer 2 – Visual synthesis (on‑demand) : when a decision requires visual comparison the system generates a PNG collage containing a filmstrip, waveform and word labels, avoiding token explosion from tens of thousands of frames.

Naïve approach: 30,000 frames × 1,500 tokens = 45 M tokens of noise. video‑use: 12 KB of text + a few essential PNGs – extremely high information density.

Full pipeline: Transcribe → Pack → LLM analysis → EDL → Render → Self‑Eval, looping up to three times until the quality check passes.

Usage example

Set up https://github.com/browser-use/video-use for me. Read install.md first to install this repo, wire up ffmpeg...

After cloning and installing dependencies, the AI asks for an ElevenLabs API key, then processes the media folder and performs:

Inventory all media assets.

Propose an editing strategy.

Wait for user confirmation.

Output edit/final.mp4.

GitHub: https://github.com/browser-use/video-use • Stars: 7,821

exo: Local AI cluster for multiple Macs/PCs

exo connects multiple Macs and PCs into a unified AI cluster, distributing model shards across devices to exceed the memory of any single GPU.

Features

Automatic discovery of machines via mDNS; no manual configuration.

RDMA over Thunderbolt 5 reduces inter‑device latency by 99 % (near‑zero in tests).

Tensor parallelism across four devices yields a 3.2× speedup.

Built on Apple MLX; MLX‑based distributed communication on macOS; GPU support on Linux is under development.

Compatible with OpenAI Chat Completions, Claude Messages, OpenAI Responses, and Ollama APIs, allowing existing AI clients to connect without code changes.

Custom model loading via a POST /models/add endpoint that accepts HuggingFace Hub model cards.

Performance data (public benchmarks, Jeff Geerling 2025)

Qwen3‑235B (8‑bit) on 4 × M3 Ultra Mac Studios with Tensor Parallel RDMA – test graphs available.

DeepSeek v3.1 671B (8‑bit) on the same hardware – test graphs available.

Kimi K2 Thinking (4‑bit native) on the same hardware – test graphs available.

Installation

macOS (requires Nix) nix run . # exo Manual macOS install

git clone https://github.com/exo-explore/exo
cd exo/dashboard && npm install && npm run build && cd ..
uv run exo

After starting, open http://localhost:52415 to view the dashboard.

Linux (CPU only, GPU support in development)

git clone https://github.com/exo-explore/exo
cd exo/dashboard && npm install && npm run build && cd ..
uv run exo

Enable RDMA on macOS 26.2+

rdma_ctl enable

GitHub: https://github.com/exo-explore/exo • Stars: 44,781

ai‑hedge‑fund: Multi‑agent stock analysis simulator

ai‑hedge‑fund implements 14 investment‑master agents (e.g., Warren Buffett, Charlie Munger, Ben Graham, Peter Lynch, Cathie Wood, Michael Burry, Bill Ackman) that each analyze the same ticker. Five analysis agents (Valuation, Sentiment, Fundamentals, Technical, Risk Manager) feed their outputs to a Portfolio Manager, which aggregates the results into a simulated hedge‑fund decision.

Capabilities

Multi‑angle stock analysis across value, growth, sentiment, technical and fundamentals dimensions.

Built‑in backtester accepting --start-date and --end-date arguments to replay historical data.

Command‑line interface for automation and a web UI for visual inspection.

Local LLM support via the --ollama flag, avoiding external API calls.

Installation and usage

git clone https://github.com/virattt/ai-hedge-fund.git
cd ai-hedge-fund
poetry install
# Analyze specific tickers
poetry run python src/main.py --ticker AAPL,MSFT,NVDA
# Use local Ollama
poetry run python src/main.py --ticker AAPL,MSFT,NVDA --ollama
# Run backtest
poetry run python src/backtester.py --ticker AAPL,MSFT,NVDA

The web UI resides in the app/ directory with detailed instructions.

GitHub: https://github.com/virattt/ai-hedge-fund • Stars: 58,972