Extracting data lineage
from source code.

The movement of data across systems — which source it came from, which code moved it, what happened on the way, where it landed. Coarsest: one system feeds another. Finest: a single field traced through every transformation between them.

The code never declares it. One transformation is spread across functions, files and imports. A destination may exist only at run time. The same logic can be written a dozen ways, so no fixed set of rules holds. It takes understanding what the code does, not matching its shape.How to extract column-level lineage from application code →

Regulators require it. BCBS 239 obliges systemically important banks to trace risk data end to end, at column level. Ten years on, the Basel Committee assessed all 31 global systemically important banks and reported that further work was needed at every one of them to attain or sustain full compliance. The requirement is settled; the means of meeting it is not.Basel Committee progress report, November 2023 ↗

ADL is an automated system that extracts data lineage from high-level programming languages. It reads the code and returns the lineage in it, at the finest grain the code holds, with anything unresolvable marked for review rather than guessed. Minimal human intervention by design.

Python today. Java and Scala next, then analytical languages — starting with SAS, still heavily used in banking. SQL, XML and mainframe languages such as COBOL are out of scope.

LLM-first. Parser-first is high maintenance and still cannot cover every style of code it meets.

No. The lineage lives in the code, not in the data — ADL extracts the footprint the code leaves behind, so your tables are never read. Supplying database metadata, which table holds which columns, sharpens column cascading; rows are never needed.

Never used for training or refinement — ADL trains nothing; it runs on closed-source models as they ship. Never stored: your code goes through the pipeline, the lineage comes out, the rest is deleted. You bring your own LLM API key, so model calls fall under your agreement with that provider, not ours. Happy to sign an NDA before you send anything.

All the major providers — Anthropic, Google, OpenAI — through their own API or a cloud platform you already run on, such as Azure or Bedrock. It needs a frontier model rather than a small one: Opus-class today, and the equivalent top tier elsewhere. Where the model runs often matters more than which one it is — it can sit inside your own tenancy, under agreements you already hold.

Two parts, kept separate. The model cost is yours — your own API key, paid to your provider directly, with nothing added by us. Then a licence for the software, kept deliberately low while we are early. What a run itself costs depends on the code, not the line count.

A software licence, not services. ADL is orchestration around frontier models — we train nothing of our own; the product is knowing how to drive them at this problem. You run it yourself; we are not a team you retain.

Lineage out of Python, delivered in a format you define. In stealth we run it with you: you bring the LLM API keys and name the output format, we point ADL at your codebase — or at an open-source repository, if you would rather not share code yet.