Answers document questions with verifiable programs that cite exact source words
The Hook
“Today, we're announcing our first neurosymbolic model that can actually handle real work with data.”
Feature Teaser
“Trained specifically for data extraction, Search Reader is tens of times cheaper than even the most budget friendly LLMs, with performance closing in on frontier models.”
Problem Agitation
“It isn't just better, it is a different type of AI model. Here's how it works.”
Product Reveal
“Frontier LLMs were trained to chat. Ask one about your data and it writes an answer that sounds right, whether it's in the data or not. So a human still has to check it. Reader was trained for exact answers. It points at your data and writes a formula, like in Excel, to get the answer.”
Feature Teaser
“It is built to dramatically reduce cost and latency...”
Feature Teaser
“...within tasks like finding facts and contradictions in large documents, building an enterprise search system, curating context within coding agents, or figuring out how many shares you keep after leaving Anthropic.”
Product Reveal
“Reader does this by taking the data heavy lookups, while the LLM handles workflow steps that are open ended.”
Feature Teaser
“It can even answer questions over long documents on its own, like finding a clause in a contract or a due date in an invoice.”
Feature Teaser
“The improvements are clear if you see them side by side. Ask Search Reader a question about your data just like you would an LLM. Get the answer back near instantly. Meanwhile, LLMs take many times longer to finish responding. Look at how the LLM searches through the data, file by file, then generates its answers token by token, which is great for a natural conversation, but slow and costly for computers. Or the LLM writes a script, runs it, reads the output, and fixes it. A single lookup takes several model calls. Reader writes one formula, like in Excel. It points at the words it needs, and your code runs it.”
Call to Action
“This represents a fundamental shift in how AI agents work with data. Join the waitlist at platform.xerj.ai”
Synthesizable in Remotion · key components of the neurosymbolic-benchmark-showdown template.
Dynamic rounded yellow rectangle expanding over exact text spans within document mockups.
Use Remotion interpolate on scaleX and opacity from 0 to 1 with spring physics (damping: 15, stiffness: 120), anchored at left.
Side-by-side numerical clock counter showing elapsed latency in hundredths of a second.
Render formatted frame-based numbers using useCurrentFrame() converted to seconds, clamping when final answer condition triggers.
Floating dark pill box displaying parsed Excel-style neurosymbolic syntax with highlighted argument references.
Sequence a slide-in translateY spring followed by child span text color toggles matching bounding box coordinates.
Purpose. Disrupt standard LLM expectations with cost/speed ratio.
Execution. High-contrast stat tables with 1000x cost disparity callout.
Purpose. Explain why generative LLMs hallucinate on data while neurosymbolic models calculate.
Execution. Chat failure demo juxtaposed against Excel-like formula generation.
Purpose. Provide indisputable empirical proof of speed advantage.
Execution. Side-by-side split screen with running digital stopwatch.
Purpose. Direct technical users to waitlist with final manifesto line.
Execution. Minimalist centered card and URL display with tagline.