Acropolis · First Cohort · 2026

Edge-AI Lab

I'm starting a small applied-AI research lab at Acropolis and looking for a few students to be its first cohort. You work in a team of two or three on one real project, mentored by me week to week. It ends in a working demo you can show people — not a report you file.

Where
Acropolis, Indore
Cohort
4–6 students, two teams
Kit
Your own laptop or phone
Ends with
A working demo

What we build

We build edge AI — intelligence that runs on the device in front of you instead of leaning on a distant cloud. The first constraint is deliberately hard: it must keep working when the internet disappears. It responds in real time, and the data never leaves the device.

Switch on airplane mode. It still works.That reveal is the point.

Both projects start on your own laptop or phone. Nothing to buy, nothing to wait for. Once the software works, we port it to a dedicated edge device next semester.

The two projects

Pick one. Each runs as a small team for the term.

01

Offline Hinglish voice Q&A

A thing you talk to in Hinglish that talks back. Speech in, a small language model, speech out — all on the device, scoped to one useful set of campus questions.

  • Speech-to-text
  • Small LLMs
  • Text-to-speech
The question

How useful can a fully local conversational system become on commodity hardware?

The hard part

Latency. A laggy assistant feels broken. Making the round-trip feel conversational, handling Hinglish cleanly, and grounding answers in real information so it doesn't make things up.

Done means

Ask it something with the device offline and get a correct spoken answer, fast.

02

Offline computer vision

One useful campus task from a camera, running entirely on-device. Live occupancy — how full a space is, counted in real time.

  • Object detection
  • Tracking
  • On-device runtimes
The question

How reliably can a small local vision system perceive a changing physical environment?

The hard part

Accuracy under real conditions. Campus lighting, crowding, and camera angles wreck a model that looked fine on clean test images. Counts have to stay steady instead of flickering.

Done means

Point a camera at a real scene and get a reliable live count with the network switched off.

Build, measure, learn.We don't only ask whether these work. We measure how fast, reliable, consistent and resource-efficient they are — and document what breaks.

The bigger picture

These two projects are the first experiments in a larger program. Edge Lab is organised around four questions. We run two this cohort; the rest are where it's headed — and strong contributors get first access to them.

Local

How much intelligence fits in the machine in your bag?

  • Offline Hinglish voice Q&A · this cohort
  • Private on-device memory
  • Edge-or-cloud router

Dependable

Can probabilistic systems behave predictably?

  • The reliable small model
  • Verify before you answer
  • The same prompt, 100 times
  • Knowing when not to know

Physical

How does intelligence understand the world around it?

  • Offline computer vision · this cohort
  • A safety camera that doesn't record you
  • Campus digital twin
  • Edge change detection

Distributed

What if intelligence is many small systems, not one giant model?

  • A swarm of tiny models
  • Edge / cloud routing
  • Distributed sensors

Campus is the proving ground. Beyond it: catching safety issues in a real workplace, a private assistant that never phones home, geospatial change spotted at the sensor. Strong teams push toward that.

Who you'd work with

Me, directly, every week. I spent about a decade at Apple working on Maps — large-scale mapping and machine learning — was a founding engineer at a startup that was later acquired, and studied operations research at Columbia. What I care about is how you build and think, not your marks.

More about me: bvaibhav.info/about · LinkedIn

What you get

  • Direct, weekly mentorship from me.
  • A real demo that's yours, and goes on your portfolio.
  • A small, hand-picked cohort. This is the first one, so you'll help shape how the lab runs.
  • First access, if you're strong, to the next projects in the program — a router, a campus digital twin, private on-device memory.

Who I'm looking for

People who build things because they want to, and who find hard problems interesting rather than annoying. You don't need prior experience or top marks. If you've made something on your own — anything — I want to hear about it.

How to apply

Tell me two things. No resume.

  1. Something you built or figured out on your own — and what the hardest part was. What did you try before it worked?
  2. Which of the two projects you'd pick, and why that one. Tell me the first thing you'd try to build, and the first thing you think would break.

If you have a GitHub, a video, or anything you've made, drop a link. Optional, but it helps.

Put Edge Lab — Your Name — Voice / Vision / Either in the subject line, so I can find and sort them.

  1. 1Send the two answers below.
  2. 2Shortlisted applicants get a small hands-on task to try over a few days.
  3. 3A short conversation with me. That's the whole process.
Email your answers →

Write to [email protected].