hi, i'm mannas 👋
i'm a curious engineer who likes making things: software, half-baked research, words on the internet, and a slowly growing shelf of 3d prints. right now i spend most days building small tools that make other builders faster.
when i'm not writing code i'm usually writing sentences, taking things apart to see how they work, or coaxing my printer through yet another nine-hour print that will fail at hour eight.
find me on twitter, linkedin, and github, or just email me.
psst, there's a paper plane and a hidden playlist somewhere on this page ♪
things i've built
glim · a drop-in ai guide for next.js apps: a glowing character that answers questions in-product and points at the exact ui element
vidura · a local counselor that reads your claude code sessions and tells you the truth about your friction
lerobot-lint · lints robot-training datasets for dead joints, frozen cameras, and encoder bugs before you burn gpu-hours on bad data
legal-bge-in · a bge embedding model fine-tuned on indian supreme court and delhi hc judgments, so a lawyer's question lands near the case that actually answers it
tetris-world-model · a tetris agent that learns to play entirely inside its own imagination: it builds a world model of the game, then trains by dreaming instead of touching the real one
places i've worked
wander product engineer ii
dec 2024 to jul 2026full stack across wander's entire surface area: website, app, and wanderos. led a redesign sprint from first mock to production, drove performance work that made pages feel instant, architected pms integrations so properties sync themselves, and squashed a small graveyard of bugs along the way.
freelancing web apps, scripts, and odd jobs for anyone who'd pay
before thatlearned: how to scope work, talk to clients like a human, and that shipped-and-imperfect beats perfect-and-pending every single time.
research (yes, just the one)
boli: a dataset for understanding stuttering experience and analyzing stuttered speech arxiv preprint · 2025
a multi-lingual stuttered speech dataset built with and for people who stutter in india: read and spontaneous speech, five stutter types annotated, released open access. months of work with three great co-authors, and i'm still weirdly proud of it.
writing
some things i've written down:
i fine-tuned an embedding model on indian case law and it actually worked jul 2026