Xinyang Wu

Software Engineer focused on AI Infrastructure

ETH Zürich MSc. Building grounded LLM assistants, production Python/FastAPI services, retrieval and evaluation systems, data pipelines, and Go/Kubernetes infrastructure.

Current engineering focus: reliable AI services, backend systems, cloud-native delivery, observability, and production evaluation workflows.

View Projects Download Resume GitHub LinkedIn Email

For referrals / recruiters

Target roles

  • Software Engineer, AI Infrastructure
  • Software Engineer, Backend Systems
  • Software Engineer, Cloud / Platform
  • Software Engineer, Search / Ranking / ML Infrastructure

Preferred locations

Zurich, Munich, London, Dublin, US. Open to relocation.

Best evidence

  • Grounded Documentation and Evaluation Assistant with deterministic evidence routing and human-in-the-loop golden sets
  • High-throughput FastAPI data platform
  • Trust & Safety ML prototypes with CI-enforced evaluation (Review Router, Abuse Signals)
  • Distributed multi-GPU data pipeline at Bosch
  • Current Go/Kubernetes infrastructure project

Featured projects

Featured system

Documentation & Evaluation Assistant Platform

Grounded LLM assistant for product documentation and evaluation workflows, built with routed retrieval, NDJSON streaming, source evidence, and reusable frontend drawers.

  • Built a FastAPI-based assistant service over generated product documentation and curated external knowledge corpora.
  • Implemented question routing across product, science, hybrid, troubleshooting, surface-chemistry, and comparison intents.
  • Designed retrieval policy with actionable vs. background evidence pools, transferability filters, and source-coverage guardrails.
  • Streamed token and completion events as NDJSON with resolved source metadata for frontend grounding display.
  • Reused the same assistant stream inside the Kinetics and Affinity Evaluation Assistant with goal, step, and data-selection context.
PythonFastAPITypeScriptLLMRAGNDJSON StreamingRetrieval PolicyCI/CD
Case Study Architecture
Project 2 Current build

Go/Kubernetes Service Platform

A production-style microservice deployed on Kubernetes with Helm, probes, HPA, ConfigMaps, Secrets, and CI/CD.

  • Built a Go REST/gRPC service with clean architecture and structured logging.
  • Containerized the service with a multi-stage non-root Docker image.
  • Deployed to Kubernetes using Helm with dev/prod values overlays.
  • Added liveness/readiness probes, resource requests/limits, ConfigMap/Secret separation.
  • Configured HorizontalPodAutoscaler with metrics-server.
  • Added GitHub Actions CI for tests, linting, Docker build, and Helm chart validation.
GoKubernetesDockerHelmGitHub ActionsHPAObservability
Case Study
Project 3

Review Router

Content-policy router that allocates a fixed human-review budget across abuse categories, treated as an operating-point and queueing problem.

  • Ran exact operating-point analysis over 63,978 scored Jigsaw test rows, surfacing precision ceilings of 84% (toxic) vs. 14% (threat) at 50% recall / 1% FPR.
  • Routed rare high-harm labels to human review by policy under a fixed reviewer budget.
  • Cut p50 time-to-action by 65% against an equal-capacity FIFO baseline while tracking p90/p99.
  • Quantified a 1.74x severe-toxic volume overshoot from train-test prevalence shift that ROC-AUC missed.
  • Enforced per-label precision, clearance, and determinism floors in CI.
PythonML EvaluationOperating PointsQueue SimulationCI Gates
GitHub
Project 4

Abuse Signals

Account-level abuse-detection prototype combining SQL behavioral signals, deterministic rules, and supervised classifiers.

  • Engineered SQL behavioral signals over 20K synthetic accounts.
  • Combined deterministic rules with supervised classifiers for throttle/suspend decisions.
  • Reached PR-AUC 1.00 with throttle/suspend precision of 1.00/1.00 and recall of 0.96/0.66 on held-out synthetic data.
  • Enforced behavior with a 12-test CI gate.
PythonSQLscikit-learnDeterministic RulesCI Gates
GitHub
Project 5

High-Throughput Time-Series Data Platform

FastAPI backend for biosensor time-series ingestion, indexing, and concurrent internal serving.

  • Designed and optimized a FastAPI data ingestion and processing platform.
  • Implemented storage-layer indexing to reduce query latency by 10%.
  • Supported concurrent serving for 30 internal users.
  • Built deployment automation with GitLab CI and Docker.
FastAPIPythonTime-SeriesIndexingDockerGitLab CI
Case Study
Project 6

Distributed Dataset Versioning & CV Pipeline

Reproducible data tooling on a multi-GPU Linux cluster for large-scale computer vision workloads.

  • Built dataset versioning and automated data-pipeline tooling.
  • Reduced model-training data-preparation time by 50%.
  • Improved reproducibility for downstream computer-vision stacks.
  • Optimized GPU memory utilization for 3D video reconstruction.
  • Work productized into first-author patent US 19/402,148.
PythonPyTorchLinuxMulti-GPUDataset VersioningComputer Vision
Case Study
Project 7

AI Reading Partner

Chrome extension using Gemini API, DOM extraction, backend orchestration, and live UI.

  • Built a multimodal agentic Chrome extension for a browser AI challenge.
  • Implemented DOM extraction, backend orchestration, and live UI updates.
  • Demonstrates compact end-to-end GenAI product delivery.
Chrome ExtensionGemini APITypeScriptGenAIRAG
Demo

Engineering Focus

AI Infrastructure

LLM services, RAG-style grounding, question routing, retrieval policy, source coverage, evaluation harnesses, golden sets, human-review workflows, model validation gates.

Backend Systems

FastAPI, REST APIs, async and NDJSON streaming, session management, indexing, query optimization, SQL, data ingestion.

Cloud Native / DevOps

Docker, GitHub Actions, GitLab CI, Kubernetes, Helm, HPA, ConfigMaps, Secrets, service health checks.

Observability

OpenTelemetry, OTEL Collector, Datadog, RED dashboards, p95/p99 latency, structured logs, runbooks.

Languages

PythonGoTypeScriptSQLJavaCBashMATLAB
View full engineering focus

Experience

Miltenyi Biotec - Software Engineer

Zurich, Switzerland

Sept 2024 - Present

Production backend, grounded LLM assistants, observability, CI/CD, and internal AI tooling.

Engineering signals

Backend ownershipGrounded LLM serviceRouted retrievalEvaluation infrastructureCI/CDObservability

Highlights

  • Built FastAPI data platform for biosensor time-series.
  • Shipped documentation and evaluation assistants with deterministic evidence routing, streaming responses, and source grounding.
  • Built human-in-the-loop evaluation infrastructure: a versioned golden set enforced by regression assertions and deterministic retrieval gates in CI.
  • Hardened production reliability by converting 46 unhandled HTTP 500 error paths into typed errors across 120 call sites.
  • Added CI/CD validation gates for backend/frontend/model updates.
  • Instrumented services with OpenTelemetry and Datadog.

Bosch - Machine Learning Engineer Intern

Stuttgart, Germany

Mar 2024 - Aug 2024

Distributed data and computer-vision infrastructure on multi-GPU Linux clusters.

Engineering signals

Distributed data pipelineReproducibilityGPU clusterContainerized environments

Highlights

  • Built dataset versioning and reproducible data pipelines.
  • Optimized GPU memory for 3D video reconstruction.
  • Automated data-pipeline provisioning and container validation.
  • Work productized into first-author patent.

ETH Zurich - Research Assistant, Generative AI

Computer Vision Lab, Zurich

Mar 2023 - Sept 2023

Controllable image generation and reproducible evaluation pipelines.

Engineering signals

Fine-tuningEvaluation pipelineModel controllability

Highlights

  • Built a controllable Stable Diffusion 1.5/ControlNet image-generation pipeline, fine-tuning trainable adapters on a 24 GB RTX 3090.
  • Improved mean color-accuracy AUC by 3.97 points over prior work with reproducible training and evaluation runs.

Technical Notes

Build logs with systems-level reasoning.

Short engineering notes on service boundaries, retrieval policy, release gates, observability, and the operational details behind production AI systems.

Build log

Designing a Grounded Documentation Assistant

A build log on turning a documentation chatbot into a routed, source-grounded assistant that can also guide evaluation workflows.

FastAPILLMRAGStreamingTypeScript
Build log

A Working Map of Computer Networking

Layering and encapsulation are the whole trick — a compact map of the five-layer model with the details that matter in practice.

NetworkingTCPHTTPDNS