Sieve

Inspiration

We noticed a glaring inefficiency in modern AI infrastructure: Retrieval-Augmented Generation (RAG) is inherently bloated. Developers are pulling 15,000+ tokens of unstructured data from their vector databases and blindly throwing it at expensive LLMs like GPT-4 or Claude. This brute-force approach leads to skyrocketing API costs, sluggish latency, and a much higher risk of AI hallucinations due to context confusion. We were inspired to build a "smart filter" that sits perfectly between the database and the LLM—a system that could surgically extract only the data that matters.

What it does

Sieve is an AI infrastructure layer and MCP (Model Context Protocol) Server that acts as a real-time token shredder. When a RAG pipeline retrieves data, Sieve intercepts the payload before it ever reaches the LLM. It uses a lightning-fast, local cross-encoder model to mathematically score the semantic relevancy of every single sentence against the user's prompt.

Sieve keeps the high-scoring, highly relevant data and instantly "shreds" the hallucination-inducing noise. Instead of sending 15,000 bloated tokens to an expensive LLM, Sieve delivers a hyper-concentrated stream of just 120 pure tokens. The result? Near-instant token generation and a 99% reduction in API compute costs.

How we built it

We focused heavily on making the invisible visible. The backend logic relies on lightweight, local NLP models (cross-encoders) designed for sub-millisecond execution to ensure our intercept overhead is virtually zero (averaging 0.08s).

For the frontend, we built a visually spectacular Next.js application to serve as our interactive demo. We implemented a custom WebGL fragment shader using Three.js for our Hero background, layered with global interactive radial gradients. To explain complex RAG pipelines, we engineered a full-screen "Scrollytelling" experience using Framer Motion. As users scroll, they physically step through the pipeline, watching infinite looping data visualizations cross-fade to compare Standard RAG's bloat against Sieve's hyper-efficiency. We also integrated Recharts for live-animated data comparisons.

Challenges we ran into

Translating abstract backend infrastructure concepts (like vector retrieval and token optimization) into intuitive, high-performance UI was incredibly difficult. We initially built a static grid layout, but it didn't capture the true "flow" of data. Transitioning to a strict, scroll-driven Framer Motion sequence required complex state management and useTransform hooks.

We had to perfectly calculate opacity bands and unmount sequences using AnimatePresence to ensure our massive arrays of animated DOM elements didn't overlap, ghost, or tank the browser's framerate. Additionally, perfectly layering our WebGL shader underneath the DOM while preserving global pointer events for hover states required meticulous z-index tracking.

Accomplishments that we're proud of We are incredibly proud of the Scrollytelling pipeline. It takes a highly technical backend concept—intercepting and scoring RAG context windows—and turns it into a visceral, cinematic experience that anyone can understand in just a few seconds of scrolling. We're also extremely proud of the overall aesthetic; we successfully executed a premium, billion-dollar "technological optimism" design system using deep dark modes, glassmorphic interfaces, and our signature Lavender accent.

What we learned

We learned that the real bottleneck in modern AI isn't the LLM itself—it's the data we feed it. Optimizing the context window is the absolute most effective way to scale enterprise AI applications. Technically, we leveled up our frontend skills massively, mastering Framer Motion's advanced scroll tracking and learning how to bridge React state with complex infinite animations without causing memory leaks.

What's next for Sieve

The next step is launching Sieve as a fully distributable npm package and dedicated MCP Server that developers can drop into their existing LangChain or LlamaIndex pipelines with a single line of code. We also plan to train our own proprietary, ultra-quantized interceptor models specifically fine-tuned for specialized industries like legal tech and medical RAG, where precision context is absolutely critical.