Agentic Fund-Ops logo Agentic Fund-Ops
Agentic data engineering · Databricks

Messy fund PDFs → governed, validated, queryable data.

An AI agent parses capital-call notices, distribution notices and LP capital-account statements, extracts structured fields with an LLM, gates the model with deterministic arithmetic, and lands governed Delta tables — then serves them through Genie, dashboards and an MLflow eval. Built end-to-end on Databricks.

View the code See the dashboards

100% synthetic data · a hands-on demonstration, not a production system.

Abstract illustration: fund documents flowing through extraction nodes into a bar chart and a validation check mark

How it works

Five set-based stages over native Databricks AI Functions — the model handles judgment, deterministic SQL handles trust.

Pipeline: ingest, parse (ai_parse_document), extract (ai_query), validate (deterministic SQL), serve (Genie + dashboard)
Parseai_parse_document

Every PDF → text in one warehouse query.

Extractai_query

Structured JSON into typed columns, schema-driven.

Classifyai_classify

Doc-type routing, cross-checked 90/91.

Validatedeterministic SQL

Arithmetic reconciliation; catches what the model misses.

ServeGenie · AI/BI · MLflow

Query, visualise, and measure the output.

The data ●

Pulled from the governed fund_ops.silver tables — queried live from Databricks on this deployment, or a baked snapshot on the static mirror. Every capital-call notice, distribution notice and LP capital-account statement the pipeline processed.

Capital called by currency

Documents processed by fund

Distributions by type

LP net asset value by fund

Capital calls trend

Measure before you trust

Two native extraction strategies, scored field-by-field against 19 hand-verified gold documents. The structured ai_query approach wins decisively — and the eval harness is why we know, rather than guess.

Deterministic validation flagged 33 anomalies

Hard checks all pass; the warnings are the value — line-item, waterfall and capital-account roll-forward breaks surfaced for human review. (20 LP capital-account statements don't foot: stated closing balance ≠ opening + period activity.)

DocumentCheckSeverity

Databricks AI/BI dashboards

Two dashboards over the same governed tables — a deep Genie-scaffolded analytical view, and the pipeline's own operational overview.

Capital Calls Analysis

Scaffolded with Databricks AI/BI + Genie

Nine analytical tabs — Executive Summary, Fund Performance Rankings, Year-over-Year, Quarterly Breakdown and more — with global fund and date-range filters.

Capital Calls Analysis — Executive Summary: KPI tiles and capital-calls trend
Executive Summary
Capital Calls Analysis — Fund Performance Rankings: rankings table and top-fund bar charts
Fund Performance Rankings

Agentic Fund-Ops Overview

The pipeline's operational dashboard

KPI tiles, capital-by-currency and documents-by-fund charts, and a live validation-anomalies review table — the same numbers the validation and eval stages produce.

Agentic Fund-Ops Overview — capital called by currency and documents processed by fund
Capital by currency · documents by fund

Built with