Research Projects: Offline Student LLMs, Phone-as-Computer Dongle, AI to Double Farmers' Income

Flagship research

Frugal AI for India

Flagship research · 01

Student LLM: subject tutors that run on the edge, with no cloud and no internet

Small language models (SLMs), one per subject, that run entirely on a low-cost device a student already has. No cloud account, no data plan, no internet: the tutor works in a village classroom exactly as it does in a city.

What we are building

Subject-specific small models for school and college subjects such as mathematics, science, English and programming, instead of one large general model.
Models compressed to run on phone-class and dongle-class hardware, fully offline.
Answers grounded in the syllabus and textbooks for that subject, so the tutor teaches what the student is actually examined on.
Explanations, worked examples, practice questions and feedback in the student's own language.

Aegis launched the Student Model on Dongle challenge, to run subject-specific small language models on hardware dongles, in the presence of Mr. Jayant Chaudhary, Union Minister of State for Education.

Research questions

How small can a subject model be before it starts teaching things that are wrong?
How do we measure correctness and safety for a tutor that nobody can monitor online?
How do we update models on devices that are never connected?

Work on this with us →

Flagship research · 02

Phone-as-Computer: a dongle that turns a mobile phone into a full computing device

Most students in India own a phone, not a laptop. This project builds a dongle that uses the computing power of the student's mobile phone and lets them connect a monitor and a keyboard, giving them a real desktop for learning and coding.

What we are building

A dongle that connects the phone to a monitor, keyboard and mouse.
A desktop environment that runs on the phone's own processor: a code editor, Python and other languages, documents and a browser.
The offline Student LLM available inside that desktop as a tutor and coding assistant.
A design cheap enough for a school lab or a family to own.

Research questions

How much real development work can phone-class hardware carry, and for how long on battery?
What is the least expensive bill of materials that still gives a usable coding environment?
How should a desktop for first-time computer users be designed?

Work on this with us →

Research analysis · 03

Doubling farmers' income with AI and data science

An Indian agricultural household earned an average of ₹10,218 a month in 2018-19, and only ₹3,798 of that came from growing crops. Wages brought in more. So if crops alone had to double the income, crop earnings would have to rise about 3.7 times. No single tool does that. This analysis asks which levers AI and data science can really move, by how much, and for whom.

Research questions

Which lever gives the largest rupee gain for a small or marginal farmer, state by state?
What data exists at village level, who holds it, and is it good enough to act on?
How do we measure income gained because of AI advice, rather than a good monsoon?
Can the Student LLM approach run a farm advisor offline, on a phone, in the farmer's language?

Work on this with us →

Where AI and data could help

Income lever	What AI and data science could do
Crop productivity	Field-level advice on sowing, irrigation and fertiliser from soil, weather and satellite data. Disease identified from a phone photo.
Livestock productivity	Animal health and milk-yield records, with early warning when an animal falls ill.
Lower cost of production	Fertiliser, water and pesticide matched to what the field needs, so farmers stop paying for inputs that add nothing.
Cropping intensity	Water and weather forecasts that show whether an extra crop season is safe to plan.
Higher-value crops	Crop choice matched to soil, water and forecast demand, with the risk shown before the farmer commits.
Better prices	Mandi price forecasts: when to sell, where to sell, and whether storing the crop is worth it.
Income beyond the farm	Wages are the largest income source, so skills and job matching for the household count too.

The levers follow the Committee on Doubling Farmers' Income (chaired by Dr Ashok Dalwai, 14-volume report, 2018). The AI uses are hypotheses for this research to test, not results. Sources: NSO survey, 2019 · Dalwai Committee report · NCAER

Social good projects

Five projects students can take up today

From the Social Impact Lab on the Assessfy marketplace. These are open project briefs with milestones and AI-assessed feedback, not finished Aegis systems.

Multilingual AI Chatbot for Government Scheme Eligibility and Application Assistance

Helps citizens check whether they qualify for government welfare schemes and apply, in their own language.

SDG 16 →

Digital Skill Passport and AI Job Match Platform for Informal Workers

Gives informal workers a record of their skills and matches them to work.

Livelihoods →

Urban Flood Early-Warning and Response Platform

Warns city residents and responders before and during floods.

Climate →

Low-Cost Wetland Health Monitoring Platform

Tracks the condition of wetlands at low cost. A natural companion to SaveNadi.

Environment →

Digital Fair-Trade Marketplace Platform for Rural Indian Artisans

An online marketplace for rural artisans, designed for low-literacy users, with digital payments and fair-pricing tools.

SDG 8 →

See all social good projects →

Proposed projects

Where the flagships lead next

Five projects that build on the Student LLM and the dongle, open for students, faculty and partner organisations to take up.

Proposed

Offline coding tutor

A small code model that runs on the phone-as-computer desktop: explains error messages, suggests the next step and reviews a student's program without sending code anywhere.

Error explanations in plain language and in Indian languages.
Hints first, answers later, so the student still learns.
Works with the languages taught in the first year of college.

Proposed

Voice tutor in Indian languages

Speech in, speech out, on the device: a student asks a question aloud in Hindi, Marathi or Gujarati and hears the explanation, for learners who are not yet comfortable reading English.

On-device speech recognition and text-to-speech.
The subject models from the Student LLM project as the brain.
Testing with real classrooms for accent and noise.

Proposed

An open benchmark for student-facing small models

Before an offline tutor reaches a child, someone has to show that it is right. This project builds an open test set that scores small models on syllabus correctness, made-up answers and age-appropriate behaviour.

Question sets per subject and class, written with teachers.
Scores that a school or a ministry can read without being an AI expert.
Published openly so any model can be compared.

Proposed

Offline teacher's assistant

The same edge models, pointed at the teacher's work: lesson plans, question papers at three difficulty levels, and first-pass marking with the teacher in control of every grade.

Runs on a school laptop or the phone-as-computer dongle.
The teacher approves everything; the assistant never grades on its own.
No student data leaves the school.

Proposed

Classroom-in-a-box

One low-cost device that serves a whole classroom over local Wi-Fi, with the subject models and learning content on board, for schools with no internet at all.

Thirty students, one box, no connectivity.
Content and model updates by memory card.
Usage reports the school keeps for itself.

Work with us

Build this with us

Hardware makers, chip and device companies, education and agriculture departments, schools and foundations: we are looking for partners to pilot and manufacture. Students can join through the Aegis programs.

Research & partnerships contact

Ritin Joshi

Head – Admissions and AI Executive Education | Learning Solutions

ritin.j@aegis.edu.in · +91 77770 66810

See the systems that are live

SaveNadi (river water quality in Maharashtra), Khozo (helping missing children return home) and Aarakshak (sovereign identity verification) are described on the Research & AI for Social Good page.

Research & AI for Social Good →

Earlier projects

Video Recommendations: Recommends faculty lecture videos and related YouTube lectures for a topic and its subtopics using machine learning.
Video Indexing & Recommendations: Indexes a lecture video by the keywords in it so a learner can jump to a topic; generates a word cloud and content index.
Image-based Attendance Taking System: Marks attendance from a classroom photograph or CCTV footage using face recognition, with reports stored on a cloud server.
Face Login on mUniversity Account: Face recognition as a second level of security on the mUniversity learning platform.
Lead Management & Tracking CRM: Scores each enquiry on a five-point scale, estimates the probability of application and admission, and shows enquiry sources by channel.
Identifying Customers and Customising Marketing: Clusters similar customers from historical data and predicts what information to share with each group next.
Lead Distribution and Employee Performance Matrix: Distributes each day's leads evenly across the available team and measures individual productivity.