Fintech Real Estate Intelligence System

What happens, step by step

In plain language, in the order it happens.

1. 1

   Find the seams

   Before splitting anything, identify which parts of the monolith actually have independent reasons to change. A split along the wrong seam produces a distributed monolith, which is strictly worse than the thing you started with.

2. 2

   Move to event-driven ingestion

   Transactions arrive as events and are processed as they land, rather than being swept up in an overnight batch. This is what took reporting from hours to something usable during a working day.

3. 3

   Hold integrity at 100%

   Zero missed financial transactions is the only acceptable number when the records are money. That constraint drove the pipeline design more than throughput did.

4. 4

   Wrap the audit in language models

   RESTful services around GPT-4 and Llama for automated financial audit, collapsing roughly fifteen hours of weekly manual review to about three.

5. 5

   Route what the model should not decide

   The reliability work was mostly about knowing when not to trust the output. Low-confidence cases go to a human rather than being resolved by the model, which is the difference between a useful audit tool and an expensive liability.

6. 6

   Make the improvement measurable

   Technical debt was tracked as a named backlog of known-bad modules with agreed remediation criteria, defined before the work started, so the 40% reduction refers to something specific rather than a feeling.

Decisions, and what they cost

Every choice worth recording has a road not taken. Both are here.

• API-agnostic microservices

  instead ofContinuing to extend the monolith

  Debt was compounding faster than features shipped. The monolith was not slow to run so much as slow to change. Independent service boundaries let deploys stop being coordinated events. The cost is real: distributed tracing, more failure modes, and a network call where a function call used to be.

• Event-driven pipelines

  instead ofScheduled batch

  Batch reporting means the answer is always as old as the last run. For an investor watching a position, that lag is the product defect. Events cost more in operational complexity and buy currency.

• Confidence routing to a human

  instead ofTrusting model output end to end

  In financial audit, a confidently wrong answer is worse than no answer. Knowing when not to trust the model is the actual engineering; the orchestration around it is comparatively easy.

• Define the tech-debt metric before improving it

  instead ofClaiming a reduction afterwards

  Technical debt is not directly measurable, so a percentage quoted after the fact is unfalsifiable. Agreeing a tracked backlog and remediation criteria up front is the only reason the 40% figure means anything, and it is the number I am proudest of and the hardest to defend.