AI शब्दकोश
Every AI term explained in plain English — with Indian context, official sources, and real examples. 46 terms and growing. Each term has its own page for sharing and search.
How to use this glossary
These entries are here to help readers decode the language around AI in India. Start here when a policy page, startup claim, learning path, or signal uses jargon that changes the meaning of what you are reading.
AI systems that can autonomously plan and execute multi-step tasks — booking appointments, writing code, browsing the web — with minimal human intervention.
IndiaAI Mission includes 'AI Agents' research as part of its Innovation Centre programme.
A research initiative by IIT Madras (iitm.ac.in) building open-source AI models for all 22 Indian languages — including IndicBERT, Samanantar dataset, and Vakyansh speech models.
Funded partly by EkStep Foundation and Ministry of IT. Their models are used by startups and government apps across India.
The simulation of human intelligence in machines — enabling them to learn, reason, solve problems, and understand language. The broad field that includes Machine Learning, Deep Learning, NLP, and more.
A key technique in neural networks that allows the model to 'focus' on relevant parts of an input when making predictions. The foundation of Transformer models like GPT and Gemini.
A government-backed large language model initiative (under MeitY/IndiaAI) designed specifically for Indian languages and use cases — healthcare, agriculture, governance, and education.
Part of the IndiaAI Mission. IIT Bombay leads the technical consortium.
Bidirectional Encoder Representations from Transformers — a Google language model that understands context from both left and right in a sentence. Foundation for many Indian NLP models like IndicBERT.
The ability of AI to understand and interpret images and videos. Used in India for crop disease detection (Fasal), medical imaging (Niramai), traffic management, and Aadhaar face verification.
The maximum amount of text an LLM can 'remember' in a single conversation. GPT-4 has ~128k tokens; Gemini 1.5 Pro has 1 million tokens. Larger context = more nuanced responses.
Labelling raw data (images, text, audio) to train AI models. India has a massive data annotation industry — companies like iMerit, Appen India, and thousands of freelancers provide this globally.
India is the world's largest provider of data annotation services, employing lakhs of workers.
A subset of Machine Learning using neural networks with many layers ('deep' networks) to learn complex patterns. Powers image recognition, speech synthesis, language translation, and more.
The Indian ministry that runs Startup India, manages AI startup recognition, and coordinates with MeitY on tech policy. Key for AI startup grants and regulatory sandboxes.
Numerical representations of text, images, or data that capture their semantic meaning. Enable semantic search — finding content by meaning rather than exact words. Used in AI-powered search engines.
Training a pre-built AI model on a smaller, domain-specific dataset to specialise it. Example: fine-tuning Llama on Indian legal documents to create an Indian legal AI assistant.
A large AI model trained on massive general data, designed to be the base for many downstream tasks. GPT-4, Gemini, Claude, and Llama are foundation models.
NASSCOM's national AI upskilling platform offering free AI/ML, data science, and cloud courses. Used by over 4 million learners across India.
Visit futureskills.nasscom.in. Free for Indian learners; certifications recognised by industry.
AI that creates new content — text, images, code, audio, video. Powers ChatGPT, Gemini, Midjourney, Suno. Driving India's current AI adoption wave across all sectors.
Google DeepMind's multimodal AI model family (Ultra, Pro, Flash, Nano). Powers Google Search AI Overviews, Bard, and Google Workspace AI features in India.
When an AI model confidently generates false or fabricated information. A critical issue for high-stakes Indian use cases (legal, medical, government). Always verify AI outputs with official sources.
Configuration settings for training a model (learning rate, batch size, number of layers) that are set before training begins — not learned from data.
A multilingual BERT model trained on 12 Indian languages by AI4Bharat/IIT Madras. State-of-the-art for Indian language NLP tasks including sentiment analysis and named entity recognition.
A Rs 10,372 crore Government of India programme to build AI infrastructure through compute access, national datasets, application development, FutureSkills, startup financing, research capacity, and safe/trusted AI work.
Launched by Cabinet in March 2024. Administered by MeitY. Check indiaai.gov.in for current compute capacity, eligibility, and application windows.
Using a trained AI model to make predictions on new data. 'Inference cost' is what you pay per API call to GPT-4 or Gemini. Contrast with Training, which is building the model.
India's first domestic LLM built by Bhavish Aggarwal (Ola), supporting 20+ Indian languages. India's first AI unicorn ($1B+ valuation). Offers Krutrim Cloud for inference.
Positioned as India's sovereign alternative to OpenAI and Anthropic. HQ in Bangalore.
A neural network trained on massive text data to understand and generate human language. GPT-4, Gemini, Claude, Llama, and India's own Krutrim/Bharat GPT are LLMs.
Meta's open-source LLM family. Llama 3.1 (405B) and Llama 3.2 are widely used by Indian startups for fine-tuning on Indian languages and building private AI deployments.
A subset of AI where algorithms learn patterns from data rather than following explicit rules. The foundation of recommendation systems, fraud detection, and predictive analytics.
The Indian government ministry overseeing digital India, AI policy, IndiaAI Mission, cybersecurity, and IT industry regulation. Key body for all tech/AI policy in India.
Website: meity.gov.in. Issues AI guidelines, regulates tech platforms, oversees IndiaAI.
AI that can process and generate multiple types of data — text, images, audio, video together. GPT-4V and Gemini Pro Vision are multimodal. Especially useful where voice, images, or mixed interfaces help people use AI without relying only on typed English.
National Association of Software and Service Companies — India's IT industry body. Runs FutureSkills AI platform, publishes India AI market reports, advocates for AI policy.
India's most important tech industry lobby. Their AI reports are used by corporates, investors, and government.
AI's ability to understand, interpret, and generate human language. Powers chatbots, translators, sentiment analysis, and — critically — Indian language AI through models like IndicBERT.
A computing system inspired by the human brain — layers of interconnected 'neurons' that learn patterns from data. The building block of modern AI: from image recognition to ChatGPT.
India's premier policy think-tank. Published India's National Strategy for AI (NSAI) in 2018, which identified 5 focus sectors: agriculture, healthcare, education, smart cities, smart mobility.
Their 2021 'Responsible AI' report and 2018 NSAI are foundational Indian AI policy documents.
AI models whose code and weights are publicly available. Examples: Llama (Meta), Mistral, Falcon, and AI4Bharat models. Critical for India to build sovereign AI without dependency on foreign APIs.
When an ML model learns training data too well and fails to generalise to new data. A model that 'memorises' rather than 'learns'. Prevented by regularisation, dropout, and cross-validation.
The 'weights' in a neural network that are learned during training. GPT-4 has ~1.8 trillion parameters. More parameters generally means more capability — but also more compute cost.
The official communication arm of the Government of India. All official announcements about IndiaAI Mission, MeitY programs, and government AI initiatives are first published on pib.gov.in.
Primary source for official India AI news. University of AI monitors PIB daily.
The skill of writing instructions to LLMs to get better, more accurate outputs. A growing career path in India — 'Prompt Engineer' is now a job title at Indian IT companies and startups.
A technique combining LLMs with a search system — the model retrieves relevant documents before generating an answer. Reduces hallucinations. Widely used by Indian enterprises building internal AI tools.
Training AI models using human preferences as feedback signals. Used by OpenAI (GPT), Google (Gemini), and Anthropic (Claude) to make models safer and more helpful.
The largest publicly available parallel corpus for Indian languages — 49.7 million sentence pairs across 11 languages. Built by AI4Bharat. Foundation for many Indian translation models.
Open-source at ai4bharat.iitm.ac.in. Used by Indian government translation services and enterprises.
AI infrastructure (compute, data, models) owned and controlled domestically by a nation — not dependent on foreign companies. India's goal: sovereign AI through IndiaAI Mission, Bharat GPT, and Krutrim.
PM Modi and MeitY have emphasised AI sovereignty as a national security and economic priority.
Breaking text into smaller units (tokens) before feeding to an LLM. Pricing for AI APIs (OpenAI, Gemini) is based on token count. Indian languages often require more tokens than English.
Reusing a model trained on one task for a different (but related) task. Enables training Indian language models using global English base models — saves compute and data.
The neural network architecture behind GPT, BERT, Gemini, and most modern AI. Introduced by Google in 2017 ('Attention is All You Need'). Uses the attention mechanism to process sequences.
A database optimised for storing and searching embeddings (vector representations). Powers semantic search and RAG systems. Pinecone, Weaviate, and Qdrant are popular. Indian startups use Supabase pgvector.
The numerical values in a neural network that are adjusted during training to minimise errors. 'Open weights' models (like Llama) share their weights publicly, enabling customisation.
Missing a term? This glossary covers the most important AI terms for the Indian context.