Work · 4 case studies

The projects I'm proud of — and what they taught me.

A mix of product, applied AI, and data infrastructure. Each one has a real problem, a real strategy, and outcomes worth the whiteboard.

HelloFresh · 2022 — 2025

Product Manager

FMCG

Standardizing Global Quality Operations

Replaced paper forms and localized spreadsheets with a single quality platform used across 13 distribution centers.

The problem

HelloFresh was running quality checks across global distribution centers on paper and localized spreadsheets. No single source of truth, no standard rollout, and generic machine translation was quietly killing adoption in non-English markets.

The strategy

Context-first design

On-the-floor research in the warehouses. The UI had to work one-handed in loud, high-pressure environments. Reframed the architecture debate with regional stakeholders — a standardized core with configurable modules, not fragmented market-specific apps.

Solving the adoption silent killer

Integrated Meta's NLLB (No Language Left Behind) framework and taught it warehouse-specific terminology. Moved from literal translations to a context-aware engine, drastically increasing tool trust.

Outcomes

13 global markets on unified quality reporting
Real-time global dashboards — from reactive metric to site-level decision tool
High organic adoption from assembly-line workers previously on paper

Reflection

Adoption blockers are often hidden in the words. And stakeholder pushback is rarely about the tech — it's about a fear of losing local autonomy. Solve for the fear, and you get the buy-in.

Read the essays on Substack

HelloFresh · 2023 — 2025

Product Manager

FMCGTechnology

Automating Global Warehouse Localization with AI

Integrated Meta's NLLB translation engine into the CMS so new features rolled out to non-English markets in minutes, not weeks.

The problem

The core logistics platform was standardized, but rollout to non-English markets was consistently delayed by manual localization. Updates took weeks to propagate. Generic translation engines failed on warehouse terminology, and workers didn't trust the tools.

The strategy

The right model

Evaluated translation stacks and picked Meta's NLLB — direct high-quality translations between language pairs like German→Polish, without English as a pivot. No 'translation telephone.'

Contextual domain adaptation

Built a warehouse-specific glossary so the engine learned that a 'bin' is a storage location, not a trash can. Domain teaching, not just deployment.

Workflow automation

Wired NLLB into the CMS as a one-click localization step for product teams. Localization stopped being an org bottleneck and became a build-time concern.

Outcomes

Localization lead time: weeks → minutes
~100% reduction in external translation cost for internal docs

Reflection

True platform scalability comes from ruthlessly eliminating systemic friction before it hits the UI. This wasn't automation — it was designing the abstraction layer so engineering velocity could match operational expansion.

Read the essays on Substack

HelloFresh · 2022 — 2025

Product Manager

FMCG

Building a High-Performance L2 Support & Delivery Engine

0 → 1 L2 support and delivery framework — shielding engineers from interruption while giving leadership true capacity visibility.

The problem

The internal platform team had no structured support process. Production issues were ad-hoc, developers were pulled out of deep work constantly, sprint velocity was eroding, and there was no data on where things actually broke.

The strategy

Incident governance & automation

Defined severity levels and SLAs to separate emergencies from backlog. Rolled out OpsGenie for on-call rotation without alert fatigue. Wrote runbooks so recurring issues stopped needing a senior architect.

Standardizing delivery infrastructure

As Jira admin, redesigned the workspace so a support ticket could seamlessly become an epic when it needed a real fix. Rebuilt sprint planning to account for support bandwidth — for the first time, capacity was actually visible.

Outcomes

Sustained lift in planned-work completion per sprint
Single source of truth for incident history and delivery progress
Framework adopted as the blueprint for other platform teams

Reflection

Operational infrastructure is a product problem. When a team is constantly interrupted, it's rarely a performance issue — it's a systems failure. Design for the workflow, and you restore the team's ability to innovate.

Read the essays on Substack

Cigna · 2019 — 2020

Data Engineering Consultant

HealthcareTechnology

Incremental Ingestion at Enterprise Scale

Moved a healthcare data pipeline from daily full refreshes to a near-real-time delta model — 95% less data moved, materially lower cloud spend.

The problem

At Cigna's scale, the full-refresh model was collapsing. Syncs took so long they collided with the next run, compute costs were prohibitive for scanning billions of rows to catch <1% changes, and a network flicker at 98% meant restarting a 12-hour job.

The strategy

High-watermark strategy

Store max(updated_at) from the last successful sync; next run only queries rows where updated_at > watermark. For source systems without a reliable timestamp, led the effort to add row-level Change Data Capture.

Idempotency & upserts

Mandated every pipeline be idempotent. Merge/upsert logic so a re-run of a failed job never corrupted destination state — critical in a healthcare context where duplicate records are a hard failure.

Strategic partitioning

Partitioned the historical backfill by year/month/region, running parallel ingestion streams. Backfill time went from weeks to days.

Outcomes

~95% reduction in daily data transfer
~$4,000/month cloud cost savings
Data freshness: daily → near-real-time

Reflection

Performance is a product feature. At this scale, the infrastructure is the product — if the data is slow or expensive, everything built on top of it is inherently flawed.

Read the essays on Substack