Real engagements, already delivered.

Data for the same participant was fragmented across five REDCap projects with dual and quadruple QR-code identifiers, bilingual French/English forms, and external GeneXpert, treatment-outcome and national-programme feeds — making reliable, timely reporting almost impossible.

I built a three-stage Python pipeline (import → process → report) with an ID-validation engine that regex-checks and repairs every QR code, links screening to specimen to laboratory result, normalises bilingual fields, extracts spirometry metrics from PDFs, and integrates four external data sources — orchestrated hourly through Airflow.

Raw exports varied by machine language (English/French) and encoding (UTF-16-LE, UTF-8-SIG, windows-1252, latin-1), study IDs were malformed or hand-typed, pooled testing produced duplicate sample records, and tests spanning midnight produced invalid negative durations.

I built a high-performance parser with a BOM-plus-chardet encoding cascade and a process-pool that auto-switches to streaming above ten thousand files, then a cleaning layer that applies 200+ French→English mappings, extracts study IDs by project-specific regex, corrects datetime rollovers, and resolves pooled cartridges back to individual specimens.

Without a shared identifier, manual reconciliation of treatment records against six studies was slow, error-prone and unscalable, leaving treatment outcomes disconnected from enrolment and screening data.

I built a two-stage probabilistic matcher — exact phone-hash lookup, then a demographic fuzzy match (rapidfuzz on names, filtered by age and gender) with weighted scoring — that reconciles treatment records to the right participant and syncs the result back to the central outcome database, with national-programme target tracking on top.

Field testing generates messy, duplicate-prone data dumps spread across many forms and sites, with no automatic way to catch quality problems or track progress against per-site testing targets.

I built an end-to-end pipeline that deduplicates and uploads field dumps into three REDCap projects (thirteen forms), then runs a configurable data-quality engine that scores records on multiple dimensions, persists query lifetimes, and builds region-aware daily and monthly testing targets using a working-day calendar with Cameroon public holidays.

Hand-written and re-typed IDs caused mismatches and duplicate entries, and a distributed team generating labels in parallel risked number collisions across hundreds of sites.

I built a QR-rendering engine (reportlab/qrcode) that lays unique, site-formatted codes into print-ready PDF sheets, backed by a file-locked, multi-user auto-increment counter so concurrent operators never collide, plus a Microsoft Graph reconciliation script that matches generated codes against site usage and study data.

Tracking progress across six plans by hand was slow, and completed work was not surfaced to the team promptly, so blockers and achievements went unnoticed.

I built a Microsoft Graph automation that delta-extracts tasks across six plans, cleans them to one row per assignee, and posts a daily Microsoft Teams digest of work completed in the last 24 hours — with an emoji progress chart, an HTML summary table and proper @mentions — while keeping a SharePoint master site list in sync.