I'm a student who likes hardware a bit too much.I read datasheets for fun. It's a red flag, I know.Half of my projects still live on a breadboard.I take things apart and usually put them back.I bring a multimeter to every group project.Marketplace problems are unreasonably fun.My soldering is better than my handwriting.I have a drawer full of half-finished boards.I still get excited when an LED turns on.I'd rather debug a board than write an essay.Pokémon is honestly partly to blame for this.

Each Pi holds a slice of the transformer and passes activations to the next one over TCP. They all send heartbeats, so when one stops answering its layers get rerouted to a neighbour instead of taking the whole chain down. Nodes drop a lot more than you'd hope, so the failure handling ended up being most of the project.

A joystick turns each direction into its own audio tone, builds the waveform on a 5-bit DAC, and plays it out of a speaker. A second board listens on a microphone, samples at 10 kHz through an interrupt-driven ADC, strips the DC offset, buffers into a FIFO, and runs a DFT to find the loudest frequency. It only moves if the same tone shows up across several windows in a row, otherwise someone coughing nearby drives the robot into a wall.

Repository ↗

An ECG electrode and a 9-axis IMU sampled together, filtered on-device, reporting heart rate, cadence, and posture live. Finding the R-peak is easy when you're sitting still and awful when the person wearing it is bouncing up and down. It lands within 5 bpm of a reference monitor.

Repository ↗

Built at Oracle, where support engineers were losing half an hour a ticket reproducing issues and digging through old RCAs. It embeds the question, matches it against a vector index of past tickets and docs, re-ranks with a cross-encoder, and answers in-thread with citations. The re-ranker moved relevance 26%; plain vector search kept confidently returning things that were about the right topic and no help at all.

Repository ↗

A two-pass assembler — symbol table on the first pass, encoding on the second, with range checks on every PC-relative offset and immediate. Then a cycle-level simulator, and the microcode driving the datapath underneath it. It's coursework, but writing all three is the reason I can picture what a processor is doing instead of just naming the stages.

Repository ↗

Built the modules that pull customer data, meeting notes, and follow-up tasks into one place across desktop and mobile, at a startup that raised $1M from TQ Ventures. I also owned the agent pipelines the team used to scaffold features, make UI edits, and run release checks, which is the part that actually made us faster.

Pulls fundamentals from an API, builds the three statements, and drops them into an Excel DCF backplate with comparables next to it. I wrote it after building the same model by hand for roughly the twentieth stock pitch.

Repository ↗

Freight auctions price themselves with a model, and I worked on what happens when the model can't or shouldn't answer — where it hands off to fallback bids, and where margin guardrails cut in — across nine hybrid shipper workflows. Mapping the auction end to end turned up about $30M of pricing leakage and six ways the model could fail without anyone noticing. Those became the diagnostics and the launch criteria.

Shipped the Slack support agent above to four engineering teams, and set up the retrieval metrics and feedback loops that told us whether it was actually improving. Separately, specced a semantic layer on top of OCI's rule-based SQL injection detection, for queries that parse perfectly fine and still read as malicious.

C and C++ across four vehicle controllers — SPI and I²C sensor integration, plus the telemetry path, which had to stay fast enough to be useful mid-run. I added watchdog timers and structured fault logging. Nobody cares about either one until integration testing, when they're the only way to find out what broke.

Weekly review sessions and office hours for 80+ students on probability, Markov chains, and stochastic processes. I've explained stationary distributions enough times now to have a favourite way of doing it.

Founder interviews, competitive analysis, and market fit work, plus 12 investment memos on differentiation, go-to-market, and defensibility. The memos were the part I liked; you have to actually commit to an opinion in writing.

Sector research and valuation work for the fund, plus five public-market stock pitches to the investment committee.