I'm Nhan Nguyen

Software Engineer Intern

Two products off one market-data spine: a live terminal for the covered-warrant desk, and KBSV's public CW portal. The hard part was never the pricing maths — it was keeping the quote firehose correct enough that a trader trusts the screen.

• Sends differential patches over the socket rather than whole snapshots. The client has to reconcile state, which is the price paid for a fraction of the bandwidth.
• Desk terminal and public portal sit on separate VMs, so a spike in public traffic and the traders' live feed never contend for the same box.
• The data layer is its own pipeline instead of app code, so deduplication, normalisation and adjusted prices are solved once and both products read one history.

Software Engineer Intern

A financial analytics platform that answers questions against a document corpus. Almost all of the wall-clock cost turned out to be waiting on I/O rather than on the model, which changed what was worth optimising — the fix was in the plumbing, not the prompt.

• Latency came down through async request flows and cached retrieval rather than a larger model — the time was going on waiting for I/O, not on inference.
• Compressed prompts before dispatch, then checked that answer quality held: cheaper and faster is only a win if the output survives the squeeze.
• Containerised the backend and ran the test suite on every push, so the gap between my laptop and QA stopped being somewhere bugs could hide.

Software Engineer Intern

A content platform where one brief fans out into drafts for several platforms. Handing the whole job to a single model gave bland, samey output, so the pipeline was split into stages that could each be judged, and fixed, on their own.

• Routes by task rather than standardising on one model: a cheap model for extraction and classification, a strong one for the actual drafting.
• Makes quality a loop instead of a prompt — every draft is scored against external SEO signals and re-prompted until it clears the bar.
• Centralised the shared dashboard state so an edit made in one view never leaves another showing something stale, which was the bug class users actually reported.

Frontend Engineer Intern

My first internship: a B2B procurement system for Rang Dong, built for procurement staff rather than engineers. The expensive failure there is not a crash — it is a typo that quietly propagates into a budget and surfaces months later.

• Put validation inside plan creation and editing, because catching a wrong figure at entry is worth more than any amount of reporting downstream.
• With 200,000+ plans in the system, navigation was the feature — getting to the right plan mattered more than how fast it drew once you had.
• Ran the weekly budget analysis across 16 IT categories myself, which is how I learned which numbers the interface actually had to put in front of people.