Generative AI Projects for Your Portfolio: 10 on Indian Data

Key takeaways

Demand for GenAI and LLM skills grew nearly 60% in 2025, and foundit expects another 58% in 2026 (foundit Insights Tracker, January 2026).1
Proof beats paper: 40% of employers in India prefer demonstrable AI skills or certifications over degrees (NASSCOM and Indeed, May 2026).2
With 1–3 years’ experience, AmbitionBox shows Gen AI Engineers at ₹7.1–8.1 lakh a year and AI Engineers at ₹10.4–11.5 lakh (pages updated 19–21 Oct 2026).3
Data on data.gov.in is under GODL-India, which allows commercial use if you credit the provider, the source and the licence; every dataset here was checked on 22 Oct 2026.4 5
An evaluation set, citations and a cost per answer say more than a chatbot demo: build three projects, and make one of them an agent.

The 10 projects at a glance

Three retrieval projects, three agents, two in Indian languages and two that prove your work can be trusted. Pick three that fit the role you want, and finish them properly.

Project	Data from	Level	What it proves
1. A tax-help assistant that knows which Act applies	Income Tax Department	Starter	Retrieval with citations, versions
2. An RBI rules assistant that tracks updates	RBI Master Directions	Intermediate	Hybrid search, metadata, dates
3. An annual-report analyst for Indian companies	SEC filings of Infosys, Wipro, HDFC Bank, ICICI Bank	Advanced	Tables, numbers you can check
4. Ask the RBI’s card data in plain English	RBI card statistics	Intermediate	Text-to-SQL, guardrails
5. A mandi price agent that answers in Hindi	data.gov.in (AGMARKNET)	Intermediate	API tools, planning
6. Your own MCP server for air quality	data.gov.in, CPCB	Intermediate	Tool design for agents
7. Intent detection in Hindi, Tamil and Telugu	Amazon’s MASSIVE	Intermediate	Evaluation, cost, latency
8. A voice FAQ in Indian languages	Google’s FLEURS, AI4Bharat’s IndicVoices	Advanced	A speech pipeline, error rates
9. Government PDFs into clean tables	DARPG’s CPGRAMS reports	Intermediate	Structured output, validation
10. A red-team suite for your assistant	OWASP Top 10 for LLM apps	Advanced	Security, tests that run

Every link below was opened on 22 Oct 2026. Portals change: if one moves, search the publisher’s site for the dataset’s name.

What makes a GenAI project stand out?

It answers a real question from real data, proves its answers with a test set, and shows its limits, its cost and what it does when it can’t answer. A chatbot over one PDF is a weekend tutorial now; it tells an interviewer little.

RAG (retrieval-augmented generation) is a way of making a large language model answer from your documents: it finds the relevant passages first, then answers from them, with citations. An AI agent is a model that takes steps on its own, such as running a query or calling an API, to finish a task.

Interviewers look for	Show it with
Grounding	Every answer cites the passage, table or row it used
Evaluation	30–50 test questions with right answers, scored after every change
Failure handling	“I don’t know” when nothing relevant is found
Cost and speed	Cost per 100 answers and response time in the README
Tool use	A model that calls a database or an API, within limits
Shipping	A live demo, and code someone else can run

The jobs are there: in September 2026, AI/ML hiring on Naukri rose 20% on the year, and fresher AI/ML hiring rose 19%.6 New to LLMs? Our AI Engineer roadmap sets out what to learn first, and our guide to GenAI jobs in India shows what each role asks for.

Can you use this data?

Yes, with credit. Data on data.gov.in is licensed under the Government Open Data License – India,4 which lets anyone use, adapt and publish it, commercially too, if they acknowledge the provider, the source and the licence.5 The rest set their own terms:

Amazon’s MASSIVE and Google’s FLEURS are under CC BY 4.0: credit their creators.7 8
AI4Bharat’s IndicVoices is CC BY 4.0 too, once you accept its access conditions on Hugging Face.9
SEC filings are free; automated tools must declare a user agent and stay under 10 requests a second.10
RBI and Income Tax Department pages: read each site’s terms before republishing whole documents; link to them instead.
No personal data in any demo: no real PAN or Aadhaar numbers, phone numbers or customer chats.

RAG: projects 1–3

Retrieval is the skill GenAI job posts ask for most. Each of these adds one difficulty: two versions of a law, rules that change, then tables of numbers.

1. A tax-help assistant that knows which Act appliesStarter

Data: the Income Tax Department’s e-filing portal: its help pages and FAQs and its latest news. Stack: LlamaIndex or LangChain, an embedding model, FAISS or Chroma, an LLM API. Proves: citations and version handling.

Two income-tax laws are in use at once: on 12 October 2026 the portal’s news cited section 440(2) of the Income-tax Act, 2025, while its 29 September notice on returns for AY 2026–27 cited section 139 of the Income Tax Act 1961.11 A good assistant asks which year you mean.

Save 40–60 help pages as text, with each page’s title, link and date.
Chunk by heading, so each chunk holds one answer.
Answer only from what was retrieved, cite the page, and say it isn’t tax advice.
Score 30 test questions, five of them unanswerable, for correctness and citations with RAGAS.

Show it: the test table, before and after. On our program, Module 8’s third week builds a tax-help assistant on the Income Tax Department’s Acts, circulars and FAQs, on Plus and Infinity. See Module 8 week by week.

2. An RBI rules assistant that tracks updatesIntermediate

Data: the Reserve Bank of India’s Master Directions, its instructions consolidated subject by subject. Stack: hybrid search (BM25 keywords plus embeddings), a re-ranker, metadata filters. Proves: retrieval over legal text that changes.

The page lists dozens of Directions, many updated in October 2026, and one on banks’ capital for credit risk takes effect only from 1 April 2027.12

Store each paragraph of 10 commercial-bank Directions, such as Know Your Customer, with its paragraph number and “updated as on” date.
Compare recall on 30 questions: keywords, embeddings, and both with re-ranking.
Cite the paragraph, and say when a rule takes effect later.

3. An annual-report analyst for Indian companiesAdvanced

Data: the Form 20-F annual reports that Infosys, Wipro, HDFC Bank and ICICI Bank file with the US Securities and Exchange Commission, free on EDGAR; the latest four were filed between 2 June and 29 July 2026.13 Stack: Docling for parsing, Pydantic for structured output, pandas. Proves: reading tables reliably.

Extract two years of income statement and segment tables for two companies into a fixed JSON schema, and check that parts add up to totals.
Answer “how did revenue by segment change?” with the figure, its table and its page.
Let the model explain; let code do the arithmetic.

Agents and tools: projects 4–6

An agent is only as good as its tools and its limits. In each of these, every number in an answer must come from a tool, never from the model’s memory.

4. Ask the RBI’s card data in plain EnglishIntermediate

Data: the RBI’s Bankwise ATM/POS/Card Statistics, a table a month (August 2026 was the latest when we checked): by bank, cards outstanding, card payments at PoS and online, and ATM withdrawals, values in ₹ thousand. Stack: pandas, SQLite or PostgreSQL, an LLM with tool calling, LangGraph. Proves: text-to-SQL with guardrails.

Load 12 months into one tidy table: month, bank, measure, volume, value.
Give the agent one tool: a read-only SQL query with a row limit.
Show the query, the result and a chart; refuse what the table can’t answer.
Test 40 questions with known answers, comparing results, not SQL text.

Show it: a short screen recording and the test table.

5. A mandi price agent that answers in HindiIntermediate

Data: data.gov.in’s Current Daily Price of Various Commodities from Various Markets (Mandi), from the AGMARKNET portal: each day’s minimum, maximum and modal wholesale price. Its API needs a key from a data.gov.in account. Stack: requests, pandas, an LLM with tool calling. Proves: planning across tools.

Write three tools: today’s prices in a state, the markets in a district, and last month’s prices from your stored copy.
Take questions in Hindi, such as “Aaj Indore mandi mein pyaaz ka bhaav kya hai?”, and answer with the market, date and the unit the data states.
Never let the model state a price no tool returned.
Show the tool log beside each answer.

6. Your own MCP server for air qualityIntermediate

Data: data.gov.in’s Real time Air Quality Index from various locations (state, city, station, pollutant, minimum, maximum and average), and the CPCB’s AQI repository for history. Stack: the MCP Python SDK, httpx, any MCP client. Proves: designing tools others’ agents can use.

The Model Context Protocol (MCP) is an open standard for connecting AI applications to tools and data.

Expose three tools: a city’s latest readings, a state’s worst stations, a station’s history.
Describe inputs and outputs precisely: the model knows only what the description says.
Return errors a model can act on, such as “no station found; try one of these cities”.

Show it: one question answered through your server, with the call log. On our Infinity plan, Module 10 opens with building your own MCP server. See Modules 9 and 10.

Languages, documents and safety: projects 7–10

These four show range: Indian languages, messy PDFs and the security questions every AI team asks.

7. Intent detection in Hindi, Tamil and Telugu: an LLM against a small modelIntermediate

Data: Amazon’s MASSIVE: the same assistant requests in 52 languages, including Hindi, Kannada, Malayalam, Tamil and Telugu, labelled with 60 intents and 55 slot types.7 Stack: an LLM API with structured output, Transformers, PEFT. Proves: the trade-off between accuracy, cost and speed.

Take the Hindi, Tamil and Telugu test splits.
Run an LLM zero-shot and few-shot, with the 60 intents as the only allowed answers.
Fine-tune a small multilingual model with LoRA, which trains a small set of extra weights instead of the whole model.
Report accuracy by language, cost per 1,000 requests and median response time.

Show it: that table, and which model you would ship, and why.

8. A voice FAQ in Indian languagesAdvanced

Data: Google’s FLEURS (102 languages, 12 of them Indian, about 10 hours of training speech each)8 and AI4Bharat’s IndicVoices (22 Indian languages, 23.7K hours of audio).9 Stack: Whisper or an Indic speech model, your index from project 1. Proves: a speech pipeline, measured language by language.

Transcribe 200 FLEURS clips in three languages and compute each word error rate (WER).
Repeat on IndicVoices: everyday speech is harder than read speech.
Send each transcript to your project 1 assistant and reply in the same language.
Time every step.

Show it: the WER table, and two failures you can explain.

9. Government PDFs into clean tablesIntermediate

Data: the Department of Administrative Reforms and Public Grievances’ CPGRAMS Monthly Report, June 2026, with ministry-by-ministry tables: grievances received and resolved, resolution within the 21-day timeline, average disposal time and feedback.14 Stack: Docling or a vision-language model, Pydantic. Proves: extraction you can trust.

Define a schema for one table, and extract it twice: with a PDF parser and with a vision-language model.
Validate: totals must add up, and every number must appear in the PDF’s text.
Hand-check 50 cells and report how often each method was right.

Stretch: six monthly reports, and a chart of the backlog.

10. A red-team suite for your assistantAdvanced

Data: your assistant from project 1 or 2, and the OWASP Top 10 for LLM Applications, whose 2025 list starts with prompt injection and sensitive information disclosure.15 Stack: promptfoo or DeepEval, Guardrails AI or Llama Guard, GitHub Actions. Proves: you think about failure first.

Write 50 attacks in English and Hindi: instructions hidden in a document, requests for the system prompt, attempts to pull personal data.
Run them on every commit, and fail the build below your pass rate.
Fix one failure, and show the before and after.

Ship it and show it

Put each project behind a link and tell it as a story: recruiters skim the README, interviewers ask about your choices.

tax-assistant-india/
├── README.md
├── data/README.md
├── src/
│   ├── ingest.py
│   ├── retrieve.py
│   └── app.py
├── evals/
│   ├── questions.jsonl
│   └── results.md
└── requirements.txt

The README’s first line says what the project answers. Then a demo link (a Hugging Face Space works), the test results, cost and response time, the limits and how to run it; data/README.md credits each source with its licence and download date. Interviewers will ask about chunk size, model choice, cost and prompt injection: our GenAI interview questions cover them.

Free places to learn: the Hugging Face LLM Course and AI Agents Course, the PEFT documentation, and Ollama to run open models on a laptop.

Questions

What are good generative AI projects for a resume?

Projects that answer a real question from real documents and prove it: a RAG assistant with citations and a test set, an agent that uses SQL or an API, and one deployed demo. The Income Tax Department’s FAQs, the RBI’s Master Directions and data.gov.in all work well.

How many GenAI projects should I have on my resume?

Three finished ones: a RAG project, an agent and one that shows evaluation or safety. Finished, tested and deployed beats ten half-built tutorials.

Do I need a GPU for generative AI projects?

No. Most of these call a model through an API or run a small open model on a laptop with a tool such as Ollama. Fine-tuning a small model with LoRA can run on a cloud notebook’s GPU.

RAG or fine-tuning: which is better for a portfolio?

Start with RAG: most GenAI roles ask for retrieval, evaluation and tool use first. Fine-tune when you can show that prompting and retrieval weren’t enough, and measure the difference.

Can I use Indian government data in a public GitHub project?

Yes, for data under GODL-India on data.gov.in, if you credit the provider, the source and the licence. Other government sites set their own terms, so link to their documents and publish your code and results.

Sources

foundit Insights Tracker, reported in Storyboard18, 13 Jan 2026, read 22 Oct 2026.
NASSCOM and Indeed, India’s AI Talent Inflection Point, May 2026.
AmbitionBox, Gen AI Engineer (updated 21 Oct 2026) and AI Engineer (updated 19 Oct 2026) salaries in India, read 22 Oct 2026.
Open Government Data Platform India, resource page and licence notice, read 22 Oct 2026.
OpenStreetMap Wiki, Open Geospatial Data from Government of India (a summary of GODL-India), read 22 Oct 2026.
Naukri, JobSpeak, September 2026, 6 Oct 2026, read 22 Oct 2026.
Hugging Face, AmazonScience/massive dataset card, read 22 Oct 2026.
Hugging Face, google/fleurs dataset card, read 22 Oct 2026.
Hugging Face, ai4bharat/IndicVoices dataset card, read 22 Oct 2026.
US Securities and Exchange Commission, Accessing EDGAR Data, read 22 Oct 2026.
Income Tax Department, e-filing portal, Latest News (notices of 29 Sep and 12 Oct 2026), read 22 Oct 2026.
Reserve Bank of India, Master Directions, read 22 Oct 2026.
SEC EDGAR, Form 20-F filing lists of Infosys, Wipro, HDFC Bank and ICICI Bank, read 22 Oct 2026.
Department of Administrative Reforms and Public Grievances, CPGRAMS Monthly Report for Central Ministries/Departments, June 2026, published through PIB, read 22 Oct 2026.
OWASP Gen AI Security Project, OWASP Top 10 for LLM Applications 2025, read 22 Oct 2026.

Generative AI projects for your portfolio, on Indian data.