Minh Vu

• Built a self-serve CSV/Excel upload for a cash-flow forecasting product, with files of up to 10,000 rows. Before this, the principal engineer had to redo the import for every new file a customer sent. Column mapping uses fixed rules first and an LLM only when the rules fail. The data goes into an append-only PostgreSQL ledger. We used it to onboard a major customer.
• Wrote a rule-based transaction categorizer that the principal engineer used for imports. Later I built an LLM pipeline that categorizes transactions automatically.
• Built React features such as a streaming AI chat that shows the assistant's progress while it answers, and a paginated transaction table.

• Full-time from Jul 2024 to Sep 2025; back remote with flexible hours since Jun 2026. Joined as the first engineer on the AI team, which grew to 4 people.
• Built theAI backend from an empty repository(FastAPI on AWS), including the first 4 of its 10 LangGraph assessment workflows: company checks, bulk company checks, invoices and contracts.
• Added a timeout and a backup chain across LLM providers within a day, after our main provider slowed to 10-minute responses. Then I split long contracts into 5-page batches that run in parallel.
• Built a benchmark on a fixed set of 18 documents to measure accuracy, cost and speed before we changed the extraction method. On that set, field accuracy for proof-of-delivery documents went from78.9% to 100%(4 documents).
• Moved all 3 document-upload flows to direct-to-S3 uploads with presigned URLs, removing two proxy hops and up toabout 250 MBof in-process file buffering per upload batch.
• Found why some assessments showed the wrong company. The team suspected the LLM, but the cause was a cache key that ignored the company-number argument. A teammate's fix then made ID lookups match exactly.

• Wrote Logstash parsing scripts for 5 client projects at this Elastic Stack consultancy, one at a Vietnamese bank and four in Singapore. They turned raw logs into structured fields. I also wrote Filebeat configs for shipping logs.
• Built 7 Kibana visualizations on-site at the bank to track latency, uptime and errors across its web, mobile and internal systems. I also made a Kibana Canvas status board that changes color at set thresholds.

University of Passau, Germany

I completed Software Analysis, which covered dataflow analysis and abstract interpretation, and Practical Parallel Programming, which covered MPI, OpenMP, GPU programming and the polyhedral model.

FPT University, Vietnam

GPA 8.94/10 · Full scholarship · Top 10 of 900+ students