Skip to content
AG

Case studies

Work where a wrong number costs money.

Client and brand names are withheld under NDA, a rule I’ll apply to your project too. Everything else - the scale, the ownership and the outcome - is exactly as it happened.

Fortune 500 Pharmaceutical · Market Access01

I own the data layer behind how a Fortune 500 sees its contracted revenue

The problem

Market access reporting is where a pharmaceutical manufacturer finds out what it is actually being paid. Every chargeback a wholesaler claims, every rebate owed to a health plan, every gross-to-net deduction that separates list price from real revenue - it all surfaces through this reporting. Five brands, 33+ production dashboards, and the commercial teams who negotiate contracts read them as fact. The backend was spread across dozens of warehouse views, and nothing automatically confirmed that a nightly load had landed correctly. A wrong number would not announce itself. It would simply sit on a dashboard until someone priced a decision on it.

What I did

I took ownership of that backend layer and treated it as infrastructure rather than reporting. I rebuilt the fragile logic onto governed base tables so the numbers have one defensible source, re-engineered the chargeback and gross-to-net views against the warehouse fact layer instead of brittle staging tables, and wrote automated post-refresh quality control that inspects every load and raises a flag before a single dashboard is opened.

Fortune 500 Pharmaceutical · Operations02

Removed a standing weekly risk from a recurring client deliverable

The problem

A weekly deliverable that went out to the client was assembled by hand: download the vendor files, clean them in Excel, reconcile them against internal data, upload the result. Hours of it, every week, with a hard deadline. The real exposure was not the time. It was that the accuracy of a client-facing file depended on one person not making a copy-paste error under deadline pressure, and there was no mechanism that would have caught it if they had.

What I did

I replaced the whole process with a Python pipeline using secure key-pair SFTP authentication. It fetches the vendor files, validates their structure before trusting them, transforms and reconciles the data, delivers the output, and emails a quality control summary showing exactly what changed against the previous week.

Healthcare Market Research03

Made the most valuable part of a study countable for the first time

The problem

In a physician market research study, the free-text answers are where the real insight lives. They are also the part that usually gets skimmed, because coding thousands of open responses by hand is slow enough that teams quietly settle for a sample and move on. The strategy that came out of the study was therefore being shaped by the numbers that were easy to count, not by the answers that mattered most.

What I did

I designed a multi-stage LLM pipeline over PostgreSQL that reads every response, classifies and extracts what it contains, and writes it back as structured rows. A validation layer scores its own confidence and holds back anything uncertain for a human to review, so the output is auditable rather than taken on faith.

Fortune 500 Pharmaceutical · Data Migration04

Changed the foundation under a live reporting product without stopping it

The problem

A claims reporting product had to move to a completely different upstream data vendor. The dashboards on top of it were in daily use by business teams, so the migration could not come with a blackout, and it could not come with a step change in the numbers either. If figures shifted after the switch, nobody would be able to tell a genuine source difference from a migration bug, and confidence in the product would go with it.

What I did

I mapped both sources field by field and documented the lineage before touching anything, rebuilt the reporting views against the new source behind a compatibility layer so the presentation stayed stable, then ran both sources in parallel and reconciled them until every variance was explained rather than assumed.

What working together looks like

Short loops. No surprises.

  1. 01

    Free 30-minute call

    You tell me the problem. I ask the questions that reveal what is actually going on. No pitch deck, no obligation.

  2. 02

    Written proposal

    Within 48 hours you get scope, deliverables, timeline and a fixed price. If I am not the right person, I say so here.

  3. 03

    Build with visible progress

    Short feedback loops. You see working output early and often, so nothing is a surprise at the end.

  4. 04

    Handover that sticks

    Documentation, a walkthrough session, and 30 days of support. You should be able to run it without me.

Your project could be the next one here.