Cover letters that sound like you — not ChatGPT.

Dear Hiring Manager,

Operating reliable, high-throughput inference for large language models while keeping cost and latency predictable is the practical problem Aleph Alpha faces when delivering enterprise AI on European infrastructure. At Delivery Hero I led the Partner Support Triage Assistant, a retrieval-augmented system over 1.4 million historical tickets that cut median time-to-owner from 4 hours 20 minutes to 1 hour 35 minutes and reduced misrouted tickets by two-thirds, while halving inference cost through speculative decoding and dynamic batch routing. That project taught me where production risk concentrates and how to measure it before and after changes.

Most of my time there focused on making the inference and retrieval stack dependable under load. I designed and shipped a hybrid retrieval pipeline combining BM25, pgvector, and cross-encoder re-ranking, and operationalised it with an evaluation playbook that runs offline benchmarks, regression checks, and safety probes before rollout. I also integrated speculative decoding across self-hosted vLLM instances and a hosted API to control tail cost, and implemented dynamic batching to smooth p95 latency. The end-to-end work included FastAPI-based serving, PostgreSQL with pgvector for the vector store, and Kubernetes for orchestration, and it became the team standard for on-call runbooks and release gates.

Sincerely, Alex Morgan

Dear Hiring Manager,

Delivering LLM-backed, low-latency features inside an editor requires grounding generation in a customer's content graph while keeping cost and latency within tight limits. At Contentful that means retrieval and transformation must be precise and fast. At Delivery Hero I led the Partner Support Triage Assistant, a retrieval-augmented production system that routes and deduplicates roughly 12,000 weekly support tickets across 14 markets and nine languages, cutting median time-to-owner from over four hours to under two and reducing misrouted tickets by two-thirds. I designed a hybrid retrieval stack using BM25, pgvector, and cross-encoder re-ranking, and I halved inference costs through speculative decoding and dynamic batch routing, which substantially improved responsiveness for internal users.

One concrete lesson I would bring to Contentful is how to trade accuracy, latency, and cost for editor-facing features. At Zalando I replaced a legacy rules engine for size and fit recommendations with an embedding-based retrieval model, which increased add-to-cart conversion by 3.1% while also lowering inference latency. To enable near-real-time freshness I led a migration from nightly Spark batches to an online store backed by Kafka and Redis, so ranking decisions used features that were seconds fresh rather than hours. Those changes forced pragmatic choices about model complexity, offline evaluation, and release gates; I also introduced mandatory fairness checks that became part of the release pipeline.

Best regards, Alex Morgan