Redseer tracks India’s internet economy, including hundreds of private, unlisted companies an investor doing due diligence needs to understand cold.
What we wanted to find out
Could we build the underlying intelligence that lets an AI agent answer real due-diligence questions about any of these companies, and then draft consultant-grade reports on top of it?
What we built
A company-intelligence graph fed by two very different sources. First, structured metrics for unlisted Indian companies (revenue, costs, and the like) broken down by product, business unit, and location. Second, gigabytes of unstructured documents: research reports, decks, and analyst-call transcripts covering the same companies.
The hard part was the company itself as an entity. Each one has multiple products, business units, and locations, each with its own revenue and costs, all sitting inside holding structures that change over time.Grofers was a division of Zomato and is now Blinkit. The graph has to know that's all one thread. We resolved that structure, then layered the parsed documents on top so an agent has full context for any question. The reporting layer on top drafts a full Consumer Due Diligence report in about 30 minutes.
Three principles from the pressure
We had a few weeks to deliver the underlying intelligence. Specialist graph builders looked at the same problem and scoped a six-month project. We found a leaner path, and three principles came out of the pressure:
- Don’t make a model re-derive what the structured data already states. Most of a company’s shape is already structured, so ingest it deterministically and save the model for the genuine gaps.
- Freeze the durable entity and hang the changes off it as timelines. Corporate structure moves, and you can’t bake that into static edges.
- Use the surrounding entities (parent company, products, locations) to resolve the same company appearing under different names across reports. General-purpose embeddings can’t.
What we figured out
A general model plus a well-organized company graph, a context graph for a whole market, is what lets a generic LLM answer like an analyst, not a bigger model. The structured layer was the product; the model was a component.
Built through 2025, then documented, hardened, and handed over to Redseer’s team. A build-and-transfer engagement, not an ongoing managed service.