systems / regulated-credit-risk-delivery
Regulated credit-risk delivery
Shipping data systems inside a large bank since 2023
Make change without losing control.
- Period
- June 2023 – present
- My role
- Data Engineer III on a credit-risk data team, after a 2023 internship on the same team. Technical ownership, mentoring, and cross-team coordination.
- State
- Ongoing since 2023.
- 19 yearsHistorical backfill2006 to 2025, corrected in place
- 3Major features, H2 2025legacy retirement, FICO 10T for CECL, performance enhancement
- zeroDowntime through cited transitionsmanager-adopted record
- dozensSKILL.md guides written for the teamtesting, Snowflake debugging, scheduling, access, and the repetitive parts of the job
- Snowflake
- dbt
- Python
- SQL
- GitLab CI/CD
- Control-M
- DataStage
- qTest
- ServiceNow
Diagram nodes, in flow order: Requirements (business + modelers); dbt models (staging · integration · load); Validation (reusable CTE checks · unit tests); Evidence (test results · lineage · docs); Review gate (reviewed MR · approvals · weekend QA); Control-M (dependencies · run order); Production (Snowflake); Runbooks (self-serve failure handling).
What the job actually is
Credit-risk data at a bank is where "it's probably right" isn't good enough. The outputs feed risk models, regulatory reporting, and decisions about real people's credit. Every change goes through validation, reviewed merge requests, formal change control, and often a weekend QA window. I've come to like that. It forces you to make the boring parts deterministic.
The work
Historical correction at scale. A FICO model migration needed version logic backfilled across nineteen years of consumer-lending history. I partnered with senior engineers to debug the update logic, found discrepancies with targeted set-difference queries, and corrected specific records instead of rebuilding or rolling back.
Legacy retirement. In September 2025 I completed a legacy consumer-lending pipeline retirement and source transition, removing dependencies on a system headed for shutdown and keeping reporting uninterrupted through the switch.
Regulatory delivery. A FICO 10T data asset supporting CECL shipped by the September 2025 deadline. A consumer-lending performance enhancement followed in November.
Reliability. I added scheduler dependencies and run-order fixes that prevent cascade failures, promoted a scheduling workflow into production with full evidence and runbook-ready docs, and wrote the ETL overview entries so the operations team can self-serve during job failures.
Validation as a product. I collapsed a multi-layer transformation into one reusable validation query that the whole team and its validators now use. I explained recurring source-versus-target date behavior once, in writing, so it stopped blocking people.
Generative AI inside the bank
I'm a member of the bank's Data Science and Engineering Guild, the group that contributes to decisions and education on how generative AI gets implemented across the company. I bring the hands-on side: what agents can be trusted with, what they can't, and what evidence a reviewer should demand before believing either.
On my own team I wrote dozens of SKILL.md guides, one for each repetitive process in my area: testing, Snowflake debugging, scheduling, access, the things a new engineer used to learn by asking. They are written for agents and people alike, and they're why onboarding takes days instead of months.
People
I mentored a teammate through a difficult repository until he could work independently. I onboarded engineers across dbt, Snowflake, scheduling, testing, and access. I'm the person on the team who taught the others how to develop with agents without letting them run the show.
The intern summer
I took a credit-card underwriting feature from requirements through the production control gate in one internship window, despite two requirement changes. Then I built a local, references-first benefits chatbot on ChromaDB and Falcon-40B with no external internet access and presented it to hundreds of employees at one of the bank's first generative-AI events. I spent the rest of the summer helping five people deploy the code. When I came back full time, the reputation came with me.