Skip to main content

Building in the Open: The AI-Native Financial Data Foundation Series

· 3 min read
Finsight Technologies
Finsight Technologies
AI-Native Financial Data Foundations

Much of the thinking behind Finsight AI-Foundry did not start as a product. It started as a long-running research and engineering series — still ongoing, with new posts published regularly — written in public by our founder, Linxiao Ma, on Data Ninjago, under the category AI-Native Financial Data Foundation. Thirty posts and counting.

If you want to understand why AI-Foundry is designed the way it is — not just what it does — that series is the place to look, and the place to keep looking as the architecture evolves.

What the series covers​

The series works through the problem from first principles, in roughly the order the ideas matured:

1. Modelling financial products and their lifecycle​

The early posts dig into the ISDA Common Domain Model (CDM): how a financial product is defined — its nature, economic terms, payouts and schedules — and how a trade then lives: Trade, TradeState, the primitive operators that change them, and the event model that captures the full lifecycle. One post walks an end-to-end novation to show how "before state plus instruction produces after state" gives you lineage and two distinct kinds of change for free.

2. Semantics as a graph​

The middle of the series asks how financial meaning should be modelled, stored and exposed so that machines can use it — modelling semantic relationships in Neo4j, querying a financial semantic graph with Cypher, and handling the messy reality that a source field may map to a semantic concept rather than cleanly matching one.

3. AI over governed semantics — not instead of them​

The most recent posts develop the case for an agentic semantic layer: the LLM as an orchestrator over governed semantic capabilities, rather than a standalone generator of answers. That includes reading notes on tool-using LLMs, GraphRAG and graph-grounded reasoning with evidence checking, plus a retrieval pattern for connecting imprecise user language ("fixed coupon") to governed field names (FixedRate).

From blog series to AI-Foundry​

The through-line of the series is exactly the architecture behind Finsight AI-Foundry:

  • industry-aligned semantics (CDM, ontologies) as the starting foundation;
  • an institution-specific semantic layer connected to real systems and fields;
  • a graph runtime that makes meaning queryable;
  • evidence and lineage attached to every answer;
  • and AI used only where interpretation adds value — bounded, validated and governed.

An ongoing series — follow along​

This is a live, evolving body of work, not an archive: the series continues as the AI-Foundry work continues, and new posts land as new parts of the architecture are designed, prototyped and tested. If you are exploring these problems in your own institution, start with the category page and read in either direction — or start a conversation with us directly.