Background
The General Intelligence Company is building Cofounder, an AI chief of staff that helps founders run a business through agents instead of constant context switching. Their long-term vision is ambitious: a world where anyone with an idea can start a business, and where “agent-native” companies operate at global scale with minimal human overhead.
Cofounder sits across the tools a modern team already uses, including Gmail, Slack, Linear, Notion, Calendar, and GitHub. It continuously turns daily operations into usable context so agents can answer questions, draft work, and surface the next best actions.
Challenge
Cofounder lives or dies on memory. If it cannot reliably recall what happened across tools, it cannot act with confidence.
To work effectively, agents need access to everything that has already happened across a business. That includes conversations, tickets, documents, calendar notes, and attachments that arrive in unpredictable formats. This data updates constantly and often lives inside PDFs, screenshots, or forwarded files.
The team needed a system that could:
- Continuously ingest data from multiple tools every 30 minutes
- Extract text from PDFs and images using OCR
- Preserve metadata like author, source, and time
- Support agent reasoning instead of simple keyword search
Before LlamaParse, the team experimented with a managed “RAG as a service” solution. It worked at first, but quickly became limiting.
“It was too expensive, too slow, and too inflexible,” Abhi explains. “We wanted more control, but we didn’t want file parsing to become our core business.”
Building and maintaining a high-quality document parsing system in-house would have pulled engineers away from what actually differentiated Cofounder: agents that can act.
Solution
The General Intelligence Company rebuilt its memory system around a clear principle: collect everything, make it searchable, then let agents think.
Every 30 minutes, a background worker checks connected tools for new information. Messages, documents, attachments, and images flow into a centralized ingestion pipeline.
This is where LlamaParse plays a critical role.
LlamaParse handles parsing across formats, including PDFs, images, and embedded files. It extracts text, applies OCR when needed, chunks content into meaningful sections, and preserves metadata that agents later use to reason about relevance and timing.
“We’re heavy users of LlamaParse for all the file extraction and PDF parsing,” says Abhi. “It just works at the scale we need.”
After parsing, Cofounder generates embeddings for each content chunk and stores them in Chroma, creating a semantic memory layer across the entire business.
When a user asks a question, Cofounder does not rely on simple vector similarity alone. Instead, it uses a two-stage approach:
- A broad retrieval step pulls a large set of potentially relevant results from Chroma.
- An agent loads that context into a sandbox and performs deeper reasoning. It filters by time, source, and ownership, cross-references multiple tools, and synthesizes a clear answer.
Impact
By using LlamaParse as the backbone of its memory system, Cofounder gained both speed and control without taking on unnecessary infrastructure complexity.
Key outcomes include:
- Lower cost and latency compared to their previous managed RAG setup.
- Significant time savings, avoiding weeks of work to build and maintain a custom parser
- Improved agent performance, with PDFs, screenshots, and attachments fully incorporated into context
- Greater flexibility, allowing the team to tune parsing, chunking, and retrieval as the product evolves
“It probably would’ve taken us a month or more to build a worse document parser ourselves,” says Abhi. “LlamaParse let us focus on the agents instead of reinventing infrastructure.”
As Cofounder expands toward agents that run entire workflows end to end, its memory system continues to scale quietly in the background. Every email, ticket, document, and conversation becomes part of a shared, searchable understanding of the business.
That foundation brings Cofounder one step closer to its long-term goal: enabling anyone to build and run a company with agents by their side.
For early access to Cofounder, register interest here today.
