AI Research at Indian Institutes: IITs and IISc

Open a translation app and type a sentence in Kashmiri, Manipuri, or even fairly common Bhojpuri. More often than not you will get a garbled, half-nonsensical result — while the same app handles French or Spanish almost perfectly. This is not a coincidence. The big global AI companies train their translation systems mainly on text scraped from the open internet, and the internet contains enormous amounts of French and Spanish writing but comparatively little written material in most Indian languages. Fixing this gap is not something you can do by "using AI" better — asking a chatbot to translate more carefully does not create the missing data in the first place. Someone has to go out, build new datasets of Indian-language text and speech, design algorithms suited to India's many scripts and dialects, test them rigorously against real sentences, and publish the results so other scientists can check and improve them. That activity — creating new knowledge and new tools, rather than just applying existing ones — is called AI research, and a large share of it happening in India is done inside a small number of institutions: the Indian Institute of Science (IISc) in Bengaluru, and several of the Indian Institutes of Technology (IITs).

Using AI Versus Researching AI

Before going further, it is worth separating two things that get blurred together constantly.

• Using AI means applying an already-built system to solve a problem: typing a prompt into a chatbot, talking to a voice assistant, or letting a photo app sharpen a blurry picture. Someone else designed, trained, and tested the model you are using.
• AI research means building the models, algorithms, and datasets in the first place — asking "can a computer learn to do X, and how, and how well, and where exactly does it fail?" — and then answering that question through experiments precise enough for other scientists to repeat and verify.

A student who writes an app that calls an existing AI model is doing software development, which is genuinely useful. A researcher at IISc or an IIT who designs a new way for a model to learn, or who builds a dataset that did not exist before, is doing something different: producing knowledge that gets checked, published, and then reused by other researchers around the world. This chapter is about the second kind of work, and about where in India it happens.

The Research Pipeline: A Concrete Example

AI research almost always follows the same broad shape, whatever the specific problem: define the problem precisely, collect or build the data needed to study it, design and build a method, test that method numerically against real examples, and publish the result so others can scrutinise and build on it. To see this concretely rather than abstractly, consider a real, ongoing problem in Indian AI research: before a computer can even translate a sentence, it first has to figure out which language the sentence is in. This sounds trivial, but several Indian languages — Hindi, Marathi, and Nepali, for instance — are commonly written in the same script, Devanagari, so a system cannot simply look at the alphabet used and be done. It has to look at patterns of which words and letter-combinations are common in each language.

Here is a deliberately simplified version of that idea, written as a short program. It is far cruder than what an actual research lab would publish, but the underlying logic — count the evidence for each possibility, and pick the one with the most support — is genuinely the core idea behind real language-identification systems.

def identify_language(text):
    hindi_letters = set("अआइईउऊएऐओऔकखगघङचछजझञटठडढणतथदधनपफबभमयरलवशषसह")
    tamil_letters = set("அஆஇஈஉஊஎஏஐஒஓகஙசஞடணதநபமயரலவழளறன")

    hindi_count = sum(1 for ch in text if ch in hindi_letters)
    tamil_count = sum(1 for ch in text if ch in tamil_letters)

    if hindi_count > tamil_count:
        return "Hindi"
    elif tamil_count > hindi_count:
        return "Tamil"
    else:
        return "Unknown"

print(identify_language("नमस्ते"))

Trace it by hand, the way a researcher checking a colleague's code would. The string "नमस्ते" (a common Hindi greeting) is made of six characters: न, म, स, ् (a joining mark), त, and े (a vowel mark). The loop checks each one against hindi_letters. Four of them — न, म, स, and त — are base consonants that appear in the set, so hindi_count becomes 4. The joining mark and vowel mark are not base letters, so they are not counted; this is a real limitation of the simplified version, not a bug — a production system would need to handle combining marks too. None of the six characters belong to the Tamil set, so tamil_count stays 0. Since 4 is greater than 0, the function returns "Hindi", and that is exactly what print outputs. The program made a correct decision using nothing more than counting, which is the same principle — count the evidence, compare the counts — that shows up again and again in AI research, just with far larger and more statistically careful counts.

Real systems built along these lines exist. IIT Madras hosts a research group called AI4Bharat, which builds and freely releases open-source datasets, translation systems, and speech-recognition tools specifically for Indian languages, because those languages are underserved by models trained mostly on English-dominated internet text. The four-line counting idea above is a toy version of one small piece of that much larger, carefully tested body of work.

Where This Research Happens

AI research in India is not concentrated in one lab with one name. It is spread across departments and centres at several institutes, each with its own emphasis.

• IISc Bengaluru. Founded in 1909 with support from the industrialist Jamsetji Tata and land granted by the princely state of Mysore, the Indian Institute of Science is India's oldest institution built specifically for scientific research rather than only teaching. Its Department of Computer Science and Automation, together with allied departments, carries out AI research spanning learning theory, computer vision, and robotics — including work on how a robot can perceive and move through an unpredictable environment, which is a genuinely hard problem, since a real Indian street or farm field is nothing like the controlled, flat floor of a factory.
• IIT Bombay. Runs the Centre for Machine Intelligence and Data Science (CMInDS), which coordinates AI research across multiple engineering departments and trains graduate students specifically in applied machine learning methods.
• IIT Delhi. Houses the Yardi School of Artificial Intelligence, which offers a dedicated undergraduate degree in AI — unlike most IITs, where AI is still taught mainly as a set of electives inside a general Computer Science degree — alongside its faculty research programmes.
• IIT Madras. Home to the Robert Bosch Centre for Data Science and Artificial Intelligence (RBCDSAI) and to AI4Bharat, the Indian-language research group described above.
• IIT Kanpur. Set up the Mehta Family School of Data Science and Artificial Intelligence, with research spanning machine learning applied to healthcare data, cybersecurity, and infrastructure systems.

Notice that none of these institutes treats AI as a single technique. It is a toolbox — search, optimisation, statistics, and learning from data — applied to very different problems depending on the department using it: reading handwritten Devanagari, controlling a robot arm, predicting rainfall patterns, or spotting an anomaly in a medical scan.

How These Institutes Connect

Each institute is an independent hub with its own labs, its own funded projects, and its own PhD and Master's students — they are not branches of one organisation. What connects them is that they publish in the same international venues, sometimes collaborate on shared national projects, and increasingly share access to expensive computing infrastructure: the Government of India's IndiaAI Mission, approved in 2024, is specifically intended to fund large shared clusters of the specialised processors (GPUs) that AI research needs, so that a university lab does not have to buy that hardware entirely on its own.

Worked Example: Why "Accuracy" Alone Can Mislead

A large part of AI research is not building a model at all — it is rigorously checking whether the model actually works, and "works" needs a precise, numerical answer, not a vague impression. Here is the kind of calculation a researcher runs before ever writing up a result.

Suppose a research team is testing an early version of a model meant to spot a plant disease from a photograph of a leaf. This is a genuinely active area of Indian AI research, because models trained mostly on datasets of crops photographed in other countries often perform poorly on Indian crop varieties, soil colours, and lighting conditions — closing that gap requires new, India-specific data and testing, exactly the kind of work suited to a university research lab rather than a product team. Imagine the team tests their model on 100 real leaf photographs collected from Indian fields: 40 of the leaves are genuinely diseased, and 60 are genuinely healthy. Running the model on all 100 photographs gives these four outcomes:

• Diseased leaves the model correctly flagged as diseased: 32
• Diseased leaves the model wrongly called healthy: 8 (32 + 8 = 40, matching the true number of diseased leaves)
• Healthy leaves the model wrongly flagged as diseased: 6
• Healthy leaves the model correctly called healthy: 54 (6 + 54 = 60, matching the true number of healthy leaves)

The simplest measure is accuracy: the fraction of all 100 photographs the model got right, correct diseased plus correct healthy, divided by the total. That is (32 + 54) / 100 = 86 / 100 = 86%. On the surface, 86% sounds like a strong result. But a researcher digs deeper and asks two sharper questions.

Precision asks: of every leaf the model called diseased, how often was it actually right? The model called 38 leaves diseased in total (32 correctly, 6 incorrectly), so precision = 32 / 38 ≈ 84.2%.

Recall asks a different question: of every leaf that was actually diseased, how many did the model catch? There were 40 truly diseased leaves, and the model caught 32 of them, so recall = 32 / 40 = 80%.

That last number is the uncomfortable one. Recall of 80% means the model missed one in every five genuinely diseased plants, telling a farmer those plants were fine when they were not. In agriculture, a missed disease can spread through a field before the next inspection, while a false alarm (a healthy plant wrongly flagged) just costs the farmer a few minutes checking a plant that turns out to be fine. Because the cost of the two mistakes is not equal, Indian agricultural-AI researchers frequently choose to deliberately accept more false alarms — a lower precision — in exchange for catching more real cases, a higher recall, even though both models might report similar overall accuracy. This is exactly the kind of judgment call, backed by a numerical argument rather than a guess, that separates rigorous AI research from simply reporting whichever number looks best.

A Common Misconception

A widespread misunderstanding is that "AI research in India" mostly means training students to use foreign AI products more skilfully, or building an app that is just a thin wrapper around someone else's model. That is software development — genuinely useful, but it is not research, because it does not produce anything new that other scientists can independently verify. Genuine AI research produces something that did not exist before: a new algorithm with a clearly stated and testable claim about its behaviour, a new public dataset, or a rigorously measured result — like the precision and recall numbers calculated above — written up and submitted to a peer-reviewed venue, a conference or journal where other AI researchers examine the work before it counts as accepted knowledge. Researchers at IISc and the IITs regularly publish at internationally recognised AI research conferences, and it is that process of outside scrutiny, not the mere use of a fashionable technology, that defines research.

A Second Worked Example: How a Robot "Learns" Instead of Being Programmed

Robotics research at institutes like IISc often needs a robot to find a good strategy for a task no engineer can write exact rules for in advance — for example, how much force a robotic gripper should use to pick up a fragile object without either dropping it or crushing it. One family of methods used for this, called reinforcement learning, works by trial, feedback, and averaging, rather than by being given a fixed rulebook.

Here is the simplest possible version. Suppose the gripper can choose one of three pressure settings — low, medium, or high — and after each attempt it receives a numerical reward: a positive number if the grip succeeded, a negative number if it failed (the object slipped, meaning too little force, or it cracked, meaning too much). The robot keeps a running average reward for each setting and, over many attempts, increasingly favours whichever setting has the best average so far.

Trial 1: try LOW,    reward = -1   -> average(LOW)    = -1 / 1        = -1.0
Trial 2: try MEDIUM, reward = +5   -> average(MEDIUM) = 5 / 1         = 5.0
Trial 3: try HIGH,   reward = -10  -> average(HIGH)   = -10 / 1       = -10.0
Trial 4: try MEDIUM, reward = +4   -> average(MEDIUM) = (5+4) / 2     = 4.5
Trial 5: try MEDIUM, reward = +6   -> average(MEDIUM) = (5+4+6) / 3   = 5.0

After just five attempts, the arithmetic already makes the right choice obvious: MEDIUM has a running average of 5.0, comfortably ahead of LOW at -1.0 and HIGH at -10.0, so the robot's strategy shifts toward choosing MEDIUM more often. Notice what the robot was never given: nobody told it the correct pressure value in advance, and nobody wrote an equation describing the object's fragility. It found a good setting purely by trying options and averaging the numerical feedback — the same basic principle, scaled up with far more states, actions, and mathematics, that underlies reinforcement-learning research on robot manipulation. If a later, sixth attempt tried LOW again and received a reward of +1, the new average for LOW would update to (-1 + 1) / 2 = 0.0 — still far behind MEDIUM, so the robot's preference would not change, which is itself useful: the strategy correctly resists being confused by a single unusual result.

Why India-Specific Research Cannot Simply Be Imported

It is tempting to assume that once a country like the United States or China builds a strong AI model, every other country can simply use it. Several concrete reasons why that assumption breaks down for India are exactly what motivate the research described in this chapter. India has 22 scheduled languages and dozens more that are widely spoken but not scheduled, most with far less written text available online than English, French, or Mandarin, so language models trained on global internet data systematically underperform on them — this is the AI4Bharat problem from the opening of this chapter. Indian crops, soils, and field conditions differ enough from the datasets used to train most agricultural vision models that those models need to be re-tested and often rebuilt with Indian data before they can be trusted, as the leaf-disease example showed numerically. Healthcare AI trained on data from one population can perform unevenly on another population with different disease patterns, so models intended for Indian hospitals need to be validated on Indian patient data, not assumed to transfer automatically. And much of India's AI deployment — in classrooms, farms, and small businesses — has to work under real constraints on internet bandwidth and computing hardware cost, which pushes researchers toward smaller, more efficient models rather than simply importing the largest model available. None of these are reasons to distrust AI in general; they are reasons a country needs its own active research base rather than only being a consumer of research done elsewhere, which is precisely the role IISc and the IITs occupy.

Summary

AI research is the activity of creating new algorithms, datasets, and rigorously verified results — as distinct from simply using an existing AI tool. It follows a repeatable pipeline: define the problem precisely, build or gather the data needed, design and build a method, test it numerically, and publish the result for other scientists to check. In India, this work is concentrated at a handful of institutions with distinct specialisations: IISc Bengaluru (learning theory and robotics), IIT Bombay (CMInDS, applied machine learning), IIT Delhi (the Yardi School of Artificial Intelligence), IIT Madras (RBCDSAI and AI4Bharat's work on Indian languages), and IIT Kanpur (the Mehta Family School of Data Science and Artificial Intelligence, spanning healthcare, cybersecurity, and infrastructure). Evaluating a model correctly requires more than one number — accuracy, precision, and recall can tell very different stories about the same model, as the 100-leaf example showed, and choosing which mistake is more costly is itself a research judgment. Techniques like reinforcement learning let a system discover good strategies through trial, numerical feedback, and averaging rather than through a rulebook written in advance. And India needs this research specifically — not merely imported foreign models — because of its linguistic diversity, its distinct agricultural and healthcare data, and its infrastructure constraints.

Check Your Understanding

1. In your own words, explain the difference between "using AI" and "AI research," using an example different from the ones in this chapter.
2. A language-identification program is tested on 50 Tamil sentences and 50 Kannada sentences. It correctly identifies 44 of the Tamil sentences and 41 of the Kannada sentences. Calculate its overall accuracy across all 100 sentences, showing your working.
3. Why might an Indian research team deliberately choose to optimise a disease-detection model for high recall rather than high precision, even knowing it will produce more false alarms? Use the vocabulary "false negative" and "false positive" in your answer.
4. Name one institute mentioned in this chapter and one specific research focus associated with it.
5. In the robot-grip example, if a sixth trial of HIGH pressure gave a reward of +2 (instead of another failure), what would the new running average for HIGH become, and would it change which setting the robot prefers overall? Show the arithmetic.

Practice Exercises

Now it is time to practice! Complete these challenges to solidify your understanding:

• Exercise 1: Write a short program that demonstrates the core concept from this chapter. Test it with at least 3 different inputs.
• Exercise 2: Find a real-world example where ai research at indian institutes: iits and iisc is used in an Indian company (like TCS, Infosys, Flipkart, or ISRO). Write a paragraph explaining the connection.
• Exercise 3: Create a mind-map connecting ai research at indian institutes: iits and iisc to at least 3 other topics you have studied.