Michael Zhu

Senior Software Engineer with 8 years of software engineering experience, most of it on Python backends and data-heavy platforms. The last three years have leaned into production LLM and retrieval work, including a voice pipeline rework that pulled p95 from 920 ms to 410 ms under live concurrency. Day to day the work is FastAPI services on Kubernetes, with lower-level languages picked up when a rewrite or a hot path calls for it.

Prefer owning a system from the design review through on-call over handing it off at merge.

Senior Software Engineer at Alluxii (2023-07 – Present)

• Own the Python side of about 12 client deployments (FastAPI, Django, AsyncIO), with Redis and PostgreSQL underneath, Docker and Kubernetes in production, and background workers carrying the slow jobs.
• Put RAG assistants into production on LangChain, OpenAI, Pinecone, and pgvector. The corpus sits around 2.1M chunks and serves about 45K retrieval queries a week at p95 under 380 ms.
• The voice stack buckled under load. Reworked the WebSocket queue and the inference handoff so it holds steady at about 120 concurrent sessions instead of backing up.
• Took a 15-year-old.NET monolith apart into ASP.NET Core and.NET 8 services on Azure Functions. The nightly batch now runs 31% faster than the old path.
• Wrote Go workers on Gin and gRPC so embedding jobs fan out without knocking over the interactive APIs at peak.
• Built a small Rails and Sidekiq console for content ops on a learning platform. The team stopped living in spreadsheets, and publish time went from about two days to same day.
• Added Rust helpers (Tokio, Axum) on the voice hot path where GC pauses kept showing up in p99. Unglamorous work, but the spikes stopped.
• Redesigned the GraphQL and REST services behind a healthcare LMS serving about 18K monthly learners, taking API latency down roughly 23%.
• Write the RFCs when we need a shared contract, sit in design reviews, and help mid-level engineers through rollouts and on-call.
• GitHub Actions with Prometheus, Grafana, and Sentry is the default setup now. Failed deploys dropped about 18% once canaries and automatic rollback became consistent.

Software Engineer at DivIHN (2019-03 – 2023-05)

• Built AWS FinTech backends for payments, reconciliation, and vendor payouts on Lambda, S3, DynamoDB, RDS, and CloudWatch. Roughly 1.8M settlement events a month, with retries and dead-letter queues so a partial failure never double-posted.
• Django and FastAPI services on Redis Streams held about 3.2K events per minute through the busy settlement windows without duplicate postings.
• Go workers pulled from SQS and DynamoDB streams when month-end volume spiked, and payout latency stayed flat instead of climbing.
• Broke payment and media pieces out of the monolith into ASP.NET Core services on Docker and Kubernetes so they could scale on their own.
• Shipped a streaming platform running in roughly 700 locations, covering on-demand video, live streams, and secure document delivery.
• Rails storefront and catalog APIs for e-commerce clients ran through Sidekiq for the heavy jobs, which kept promotions and inventory sync off the checkout path.
• Wrote Rust parsers for bulky media manifests, shortening cold start on large batches before the data reached the streaming services.
• Owned alerting around payment and reconciliation failures. Mean time to notice a problem went from about 25 minutes to 8.
• Cut the manual deploy and queue chores down to a script, so routine patches usually went out the same day instead of sitting in a weekly batch.

Software Engineer Intern at DivIHN (2018-09 – 2019-02)

• Ran preprocessing and TensorFlow experiments on biomedical datasets in AWS.
• Wrote Ruby scripts that caught labeling mistakes before the data landed in the training set.

Master of Science in Mathematics – University of Utah (2017-09 – 2018-06)

Bachelor of Science in Applied Mathematics – University of Utah (2013-09 – 2017-06)