I’m Rounish Chandra — I build things that were meant to work, and fix the ones that weren’t. M.Tech at NIT Jalandhar, formerly hardware engineering at Microsoft. These days I move between board-level design and the AI automation that runs on top of it.
The rest is in the résumé. These are the ones where the hard part came after the demo worked.
A low-power SNN for FPGA-based detection. Digital verification, testbench automation and ONNX quantization to cut the energy cost of embedded inference — the interesting constraint was never accuracy, it was joules per frame.
A LoRA adapter trained on patch and dipole geometries so the model proposes antenna variations for a target band, with the prompt-and-select loop automated so a usable set comes out without hand-picking every frame. Then the honest part: modelled return loss against what the bench actually measured, which is where the interesting disagreements live.
A geo-aware travel planner with an A* core weighing safety, lighting, noise and construction risk, and an AI layer that says why this route beat the alternatives. Getting the pathfinding right was the easy part.
Hardware taught me the expensive version of patience: at tapeout there is no patch release, so you either understood the constraint or you didn’t. I carry that into software — I read the datasheet, I reproduce the bug before I fix it, and I’d rather ask an awkward question in week one than apologise in week six.
I like being the person who takes the boring, repeated thing and quietly automates it. Most of my freelance work starts as someone saying “we just do this manually every week.”
“Built to work. Fixed when it didn’t.” — that isn’t a tagline, it’s most of my weeks.
Requirements are usually a first draft. I'd rather spend an hour finding the assumption nobody stated than a week building around it.
Schemas, edge cases, why a constraint exists. If I leave a project, the next person should not have to reverse-engineer my reasoning.
A bug I cannot recreate is not a bug I have fixed — it is one that has agreed to wait. Hardware is unforgiving about this and it made me better at software.
If it happens weekly and a human does it by hand, it is a task waiting for a script. Usually the hard part is the exceptions, not the happy path.
Short updates, no surprises near a deadline. Bad news early is just information; bad news late is a problem.
Schema-driven Python framework decoding raw Common Platform Error Records into structured JSON — GUID-based schema registries, binary interpretation, automated error analysis.
Validation automation covering every BMC function sensor-group case, corners and edges included — then raising coverage by reproducing and fixing what it surfaced.
A small C tool reading load, temperature and memory straight off the kernel interfaces and printing them as a live terminal dashboard. Written because the ones already to hand told me less than I wanted to know.
Nature, mostly — and whatever the light happened to be doing that day. Shot on the move, over the last few years.
Open to full-time roles and freelance automation work. Fastest reply is email or WhatsApp.
A role you’re hiring for. Something at work that eats your Mondays and probably shouldn’t. A half-formed idea you want a second opinion on. Or just a hello — I read those too, and I do reply.
Opens your mail app, drafted to rounishchandra@gmail.com.
