Most agent demos stop at the model call. I build what's around it — orchestration, tool servers, retries, evaluation — the part that decides whether it survives contact with production.
|
Agent infrastructure MCP servers and orchestration frameworks that give LLMs real tools — LangGraph pipelines, Claude Code integrations, multi-step research and generation systems. |
Applied AI products End-to-end tools people actually use — video object removal and inpainting, presentation generation, asset discovery, content and media pipelines, voice and video automation. |
Fundamentals Comfortable below the framework line — implementing the underlying math by hand instead of only calling an API. |
01 video-object-remover — Click an object in a video and get back either a matte that tracks it or the plate with it removed. SAM 2 tracks the object across the clip, ProPainter reconstructs the background behind it; the track is cached by prompt, so pulling a matte and then removing the same object pays for tracking once. Ships as a desktop app, a local web app, and a CLI.
Python SAM 2 ProPainter PyTorch FFmpeg
02 propainter-delogo — Removes a static logo or watermark and rebuilds what was behind it, instead of smearing pixels over it like ffmpeg delogo. Windowed and shot-aware, so only a crop around the logo is inpainted and composited back at native resolution — every other pixel is byte-for-byte the source. 4K-friendly, audio preserved.
Python ProPainter OpenCV FFmpeg CUDA / MPS
03 clipping-mcp — MCP server that gives Claude Code 41 tools to edit video and audio, on FFmpeg + ElevenLabs: trimming, multicam podcast editing, AI voiceover, timeline projects rendered through Remotion.
Python FFmpeg MCP ElevenLabs Remotion
04 Deep Research Framework — Plugin-based multi-agent research system on LangGraph. Decomposes a query, retrieves from multiple sources, reflects on what's missing, and synthesizes a cited report, with automatic tool fallback and complexity-based model routing.
Python LangGraph Multi-agent orchestration
05 NanoBanana Shot Division — Cinematic shot-division workflow for Claude Code. Takes a creative brief, plans an adaptive DAG, and generates continuity-consistent multi-shot storyboards on fal.ai's image models.
Python Claude Code fal.ai
06 Reel Asset Engine — AI asset search engine that aggregates images, video, and articles from SearxNG, Pinterest, Unsplash, Freepik, and Pexels, then uses LLMs to score and summarize what it finds.
Next.js TypeScript LLM scoring
07 PPT Maker — Turns a prompt into a structured, styled deck in seconds — nine layout variants, tone and purpose controls, built on GPT-4.
Next.js TypeScript OpenAI Tailwind
08 Neural Network from Scratch — A neural network with no framework underneath it: forward pass, backprop, and gradient descent written by hand in NumPy, to know exactly what PyTorch is abstracting away.
Python NumPy Pandas
Building agent infrastructure and shipping applied AI products as co-founder at Boring People. Interested in tool-use, multi-agent orchestration, and the unglamorous plumbing that makes an agent reliable in production.
mail subhapattnaik@theboringpeople.in
in in.linkedin.com/in/subha-pattnaik-929488222