Acropolis · First Cohort · 2026

Edge-AI Lab

I direct a small applied-AI research lab at Acropolis in Indore. Student teams investigate what local models can do under real constraints: limited hardware, mixed-language speech and difficult camera conditions. We build working systems, measure their trade-offs and document where they fail.

Where
Acropolis, Indore
Cohort
Two teams, from October 2026
Hardware
Student laptops and Android phones
Aim
Working demos, measured trade-offs

What we investigate

We build edge AI: intelligence that runs on a local device. The first requirement is offline operation. The research question is what it takes to make that intelligence useful and dependable within the device's memory and response-time limits.

Both projects start on student laptops, with Android deployment tested as part of the work. If a phone cannot run the system well enough, we report that limit and demonstrate the laptop version. In the offline demo, audio and images are processed locally. Cloud models are used separately as evaluation baselines on data approved for that purpose.

Airplane mode makes the constraint visible.The result is a measured account of what works, how quickly, and where it breaks.

The lab is one place I investigate dependable intelligence in real systems.

The two projects

One team on each, for the term.

01

Offline Hinglish voice Q&A

A spoken Hinglish interface to one fixed set of campus information. Speech recognition, retrieval, a small language model and speech synthesis run locally.

  • 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

Recognizing mixed Hindi and English, preserving meaning, and grounding answers in the source text. Each stage adds delay; refusing an unsupported question matters as much as answering a supported one.

The target

A working offline conversation, evaluated for answer correctness, refusals and time to first spoken response. Report the hardware and the failures alongside the demo.

02

Offline vision: people counting

A camera-based count of people visible in campus scenes, processed locally without identifying individuals. Camera coverage and placement are part of the experiment.

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

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

The hard part

Backlight, crowding and occlusion cause missed detections. Tracking and smoothing may steady the count, but can also delay genuine changes. A camera cannot directly count someone outside its view.

The target

A live offline count, compared with hand-labelled campus footage. Measure count error, stability, response to movement and frame rate on the chosen device.

Build, measure, learn.Compare a baseline, change one part of the system, and measure the effect. Quality, latency and resource use belong together in the result.

The first cohort

The first cohort began in October 2026: two student teams working on campus at Acropolis, with weekly reviews. Students own the implementations, datasets and evaluation work. I help frame the questions, design the experiments and review the engineering decisions.

The cohort is working toward a December demonstration and technical write-ups with measured results, limitations and contributor credits. Results have not yet been published. Future student calls will appear here.

Bring a relevant problem

Have a problem where connectivity, privacy or response time limits what AI can do? I welcome conversations with practitioners who can bring real constraints and with technical peers who can challenge our methods.

For a suitable problem, we can discuss a focused, paid feasibility study. Commercial engagements are scoped and funded separately from the educational lab; students who join delivery work are hired and paid for that engagement.

Get in touch about a problem → · Other ways to collaborate