How We Built an Investment Research Agent (Kimi K2 + GPT-5 + Claude)

In venture capital and private equity, speed and depth of insight are the currency of the realm. The firm that completes its due diligence automation faster and surfaces the most critical insights wins the deal. Yet, most investment teams are still buried in a fundamentally manual financial research workflow: hours spent hunting down filings, parsing PDFs, and manually synthesizing data before a single strategic thought can be formed.

This is a classic problem that demands more than just a simple script; it requires a sophisticated, multi-talented AI. This is a task for a deal research agent.

We recently built a custom agent for a client to solve this exact problem. This post breaks down the architecture of our investment research automation system, showing how we orchestrated a team of best-in-class models (leveraging the distinct strengths of models like Kimi K2, Claude, and next-gen engines like GPT-5) to transform a week of manual work into a few minutes of autonomous analysis.

The Architecture: A Five-Step Agentic Workflow

A robust investment research agent isn't a single AI; it's a coordinated pipeline of specialized tasks. Our approach follows a clear, auditable flow from raw input to polished output.

Step 1: Ingesting the Target List

The process begins with a simple list of target companies. This could be a CSV export, a list of URLs, or an integration with a CRM. The agent’s first task is to treat this list as a queue of jobs to be executed.

Step 2: Autonomous Document Fetching

For each company, the agent must gather the raw materials for its analysis. This is a critical step that goes far beyond a simple web search.

• The Challenge: The required documents are often scattered across disparate, unstructured sources: SEC filings (10-Ks, 10-Qs), recent press releases, industry reports, and proprietary data room files.

• The Agent's Action: The agent uses a variety of tools to fetch and ingest this data. It can access public databases like EDGAR, connect to news APIs, or securely access a provided virtual data room. The key here is using a model with a massive context window, like Kimi K2, to ingest entire documents—sometimes hundreds of pages long—without truncation or loss of detail.

Step 3: Retrieval-Augmented Generation (RAG)

With the documents gathered, the agent needs to extract the most relevant facts. This is where Retrieval-Augmented Generation (RAG) comes in.

• The Challenge: We don't want the AI to "hallucinate" or provide generic summaries. We need verifiable facts grounded in the source material.

• The Agent's Action: The agent indexes all fetched documents into a secure, private vector database. When given a research directive (e.g., "Analyze revenue trends and identify key risks mentioned in the last two 10-Ks"), the RAG process retrieves the specific text snippets that contain the answer before the generative model synthesizes them. This ensures every claim in the final report is attributable to a source document.

Step 4: Multi-Model Synthesis

This is the intellectual core of the workflow. Once the facts are retrieved, a powerful reasoning engine must synthesize them into a coherent narrative.

• The Challenge: The task requires nuance. It needs to compare data points from different documents, identify contradictions, and summarize complex financial language into clear, concise insights. A simple script or a less powerful model would fail here. This is why a one-size-fits-all approach using simple RPA vs AI Agents is insufficient for high-stakes financial analysis.

• The Agent's Action: We use an orchestrator to pass the retrieved facts to a best-in-class synthesis model (like GPT-5 or Claude). This model's job is not to find information, but to reason over it. It drafts the executive summary, identifies the key risks and opportunities, and structures the data into a logical format, mirroring the analytical process of a human associate.

Step 5: Formatting and Delivery

The final step is to deliver the synthesized research in a format that is immediately useful for the investment team.

• The Challenge: The output needs to be a polished, professional document, not a raw text dump.

• The Agent's Action: The agent takes the synthesized text and uses the Google Docs API to format it into a branded, perfectly structured research brief. This document, complete with headings, tables, and bullet points, is then saved to a designated team folder, and a notification is sent via Slack or email.

The Result: A Flawless, On-Demand Research Workflow

By breaking down the financial research workflow into these discrete steps and applying the best AI model for each task, we built a deal research agent that is more than just an automation—it's a force multiplier for the entire investment team.

This level of automation requires a deep focus on security and reliability. To learn more about how we ensure enterprise-grade security, you can read about our Agent Safety & Audit principles.

Ready to build a custom research agent for your team? Explore our AI Agents for Finance to get started.