I build things that show their work.

A no_std crate targets an RTOS with no heap and no OS — and Rust’s test runner needs std. Nothing in the firmware could be tested at all without the board on the desk.

Wrote the I2C driver register by register from the InvenSense datasheet — 1024-byte FIFO, config-derived packet sizing, overflow recovery — then built a mock I2C bus with transaction logging and five injectable fault modes.

Blood glucose measurement means a finger prick. A photoplethysmography signal is non-invasive and cheap — but only useful if the error is small enough that a clinician would act on it.

Built a dual-stream model over STFT and DWT features, pairing graph convolutions with cross-attention Transformers, then quantised the 1.16M-parameter network for on-device inference.

2.3 MB at FP16 and 1.2 MB at INT8 — 76 ms and 52 ms inference latency respectively.

A collections team reported recovery up 11% month on month. Before anyone funds whatever caused it, somebody has to check whether the number is real.

Rebuilt recovery from 17 raw CSVs as a reproducible pipeline — profiling, dedup and reversal rules, an as-of joined golden dataset — then wrote the same logic as production-shaped SQL, with data-quality assertions that gate publication.

A failed recurring debit in India gets a hard-capped number of NPCI retry attempts. Retry at the wrong hour and the attempt is spent on an account that was never going to have money in it.

Built an agent that decides whether, when and how to retry — a hazard model of when the payer is actually likely to be funded, run against the cap on attempts rather than a fixed schedule.

Nobody acts on an LLM’s summary of a paper they can’t verify. The blocker isn’t answer quality — it’s that nothing shows where the answer came from.

A Next.js chat workspace over a curated library of ML papers, with section-aware chunking rather than whole-paper retrieval, so evidence lands on the right part of a paper.

Project managers find out a task has slipped at the point where it’s too late to act — even though the slippage is visible in the task log days earlier.

Defined the MVP as a ranked at-risk task queue with forecast completion dates, and specified the 13-field task-log schema and the risk thresholds behind it.

Four things every NSUT student does badly through scattered channels: finding scholarships, recovering lost items, practising for assessments, and making sense of a syllabus.

Led product — turned recurring student complaints into four scoped modules, then drove the scope and sequencing calls, cutting features to protect a 36-hour deadline.