InternshipAll 60 tasks, level by level
The real graded Generative AI & Prompt Engineering tasks the portal assigns, grouped into three levels.
The portal assigns these Generative AI & Prompt Engineering tasks. Finish Level 1 before Level 2 and Level 2 before Level 3. Your mentor may mark some tasks as required for the certificate.
Level 1: Foundations (22 tasks)
| Task | What to do |
|---|---|
| Compare three AI assistants | Ask the same 5 questions (coding, maths, summary, translation, a fact) to three free AI assistants. Record answers in a table and note which was right, wrong or vague. |
| Write a prompt-engineering cheat sheet | Document role, task, context, format and constraints with one before/after example each. Commit it as README.md in a repo named genai-internship. |
| Zero-shot vs few-shot prompting | Classify 20 product reviews as positive/negative with a zero-shot prompt, then with 3 examples. Report accuracy for both and explain the difference. |
| Structured output with JSON | Write a prompt that turns 10 messy job postings into JSON (title, company, skills[], salary). Validate the JSON with a short Python or JS script. |
| Spot hallucinations | Ask an AI 10 questions about your college or city. Verify each answer from a real source and list which facts were invented. Write 3 rules to reduce this. |
| Prompt for a study plan | Build a reusable prompt that creates a 4-week study plan for any subject and level. Show outputs for two different subjects. |
| Summarise a long article | Summarise a 2,000+ word article in 3 lengths (tweet, paragraph, one page). Check every summary against the source for errors. |
| Call an LLM API | Use any free LLM API from Python or Node.js to send a prompt and print the reply. Keep the API key in an environment variable, never in the code. |
| Token and cost calculator | Write a script that estimates tokens for a text and the cost for 1,000 requests at a given price per million tokens. |
| AI image prompt gallery | Create 6 images with a free image generator for one theme. Record each prompt and what you changed between versions. |
| Responsible AI checklist | Write a one-page checklist covering privacy, bias, copyright and disclosure for using AI at work, with a real example for each point. |
| Chain-of-thought for reasoning | Give an AI 10 word problems with and without asking it to reason step by step. Compare correctness and explain when it helps. |
| Prompt library in Notion or Markdown | Collect 15 tested prompts for study, coding and career tasks, each with the output it produced and notes. |
| Compare model settings | Run the same prompt at 3 temperatures and 2 max-token limits and describe how the output changes. |
| AI for coding, used honestly | Use an AI assistant to explain, test and refactor your own code; record what it got wrong and how you checked. |
| Persona prompts | Write prompts for 4 personas (tutor, interviewer, reviewer, translator) and test each with 3 inputs. |
| Detect AI mistakes in maths and code | Give an AI 10 tricky problems, verify every answer yourself and categorise the mistakes. |
| AI policy for your college project | Write rules for using AI in a team project: allowed uses, disclosure and checking outputs. |
| Prompt patterns | Practise 6 patterns (role, few-shot, chain of thought, critique-and-revise, template, delimiters) with examples. |
| Multimodal prompts | Use a model that accepts images to describe charts and screenshots; check every claim. |
| AI for learning | Use AI to create flashcards and practice questions for one subject; verify answers from the textbook. |
| Cost comparison | Compare prices of 3 model sizes for a task and estimate monthly cost for 10,000 users. |
Level 2: Building real features (20 tasks)
| Task | What to do |
|---|---|
| Build a CLI chatbot with memory | A terminal chatbot that keeps the conversation history, has a system prompt and a /reset command. Push it with a README and sample transcript. |
| Prompt templates with variables | Build a small library of 5 prompt templates (email, summary, quiz, code review, translation) filled from variables, with tests for each. |
| Extract data from PDFs | Extract invoice number, date, items and total from 5 sample PDF invoices using an LLM, saving the results as CSV. Report any field it got wrong. |
| Build a quiz generator | Given a topic and difficulty, generate 10 MCQs with answers and explanations as JSON, then show them in a simple web page. |
Function calling / tool use | Let an LLM call 2 functions you write (weather lookup with mock data, calculator). Log every tool call and result. |
| Evaluate prompts with a test set | Create 30 test inputs with expected outputs for one task. Score 3 prompt versions automatically and pick the best with numbers. |
| Content moderation filter | Classify 50 sample comments into safe/spam/abusive with an LLM. Measure precision and recall against your own labels. |
| Streamed responses in a web app | Build a small web page that streams an LLM answer token by token from your backend instead of waiting for the full reply. |
| Resume feedback tool | Upload a resume (text or PDF) and get structured feedback: strengths, gaps and 5 rewritten bullet points. Remove personal data before sending. |
| Multilingual assistant | Answer the same question in English, Tamil and Hindi. Have a native speaker rate the quality and record the scores. |
| Mock interview bot | Build a bot that asks role-specific interview questions one at a time and scores answers with feedback. |
Document Q&A with citations | Answer questions about a PDF with page citations; show 5 correct answers and 2 refusals. |
| Batch processing with an LLM | Process 100 rows of text (classify or summarise) with retries, rate limits and a progress log. |
| Image generation workflow | Create a consistent set of 6 images for a brand using a style guide prompt and document the iterations. |
| Speech to text to summary | Transcribe a short talk with a speech model and summarise it into notes and action points. |
| Email reply assistant | Draft replies to support emails with tone options and a human approval step. |
| Code review bot | Review a pull request diff and comment on bugs and style; measure how many comments are correct. |
| Translation quality check | Translate 20 sentences into Tamil or Hindi and back; score meaning preservation. |
| Structured data extraction at scale | Extract fields from 200 short documents with schema validation and an error report. |
| Content moderation pipeline | Combine rules and an LLM to flag harmful content, with logging and appeal handling. |
Level 3: Professional work and capstone (18 tasks)
| Task | What to do |
|---|---|
Retrieval-augmented Q&A over your notes | Index 20+ pages of your own notes with embeddings, retrieve the top chunks and answer questions with citations. Show 10 questions it answers correctly. |
| College helpdesk bot | A bot that answers admission, fees and timetable questions only from a provided document and says 'I don't know' otherwise. Deploy it publicly. |
| Guardrails and prompt-injection tests | Write 15 injection attempts against your bot (e.g. 'ignore your instructions'). Add defences and show the before/after results. |
| Cost and latency dashboard | Log tokens, cost and response time for every request of your app and show daily totals in a simple dashboard. |
| Fine-tune or few-shot a classifier | Build a support-ticket classifier (billing, technical, other) and compare prompting vs a small fine-tuned or embedding-based model on 100 tickets. |
| AI meeting-notes app | Upload a meeting transcript and produce a summary, decisions, action items with owners and dates, exported as a PDF or Markdown file. |
| Capstone: deploy a GenAI product | Ship one complete GenAI app with sign-in, rate limiting, error handling and a privacy note. Submit the live URL, repo and a 3-minute demo video. |
| Write a case study | Publish a 1,000-word case study of your capstone: problem, prompts, evaluation numbers, costs and what you would do next. Share it on LinkedIn. |
| Content assistant for a small business | Generate weekly social posts from a product list with a human review step before posting. |
| Evaluate with an LLM judge | Score your app's answers with a rubric-based LLM judge and compare against your own ratings on 20 items. |
| Prompt versioning | Store prompts in code with versions and tests so a prompt change that breaks outputs fails CI. |
| Privacy-safe AI app | Add PII detection and masking before sending text to a model, with tests for names, phones and emails. |
| Demo day presentation | Present your best GenAI project in 5 minutes: problem, demo, numbers, risks and next steps. |
| Fine-tuning vs prompting study | For one task, compare prompting, few-shot and a fine-tuned small model on quality and cost. |
| Voice assistant prototype | Speech in, LLM answer, speech out, with latency measured at each step. |
| Guarded customer FAQ bot | Answer only from approved FAQs, escalate everything else, log every conversation. |
| Bias and safety testing | Test your app with inputs about gender, region and language; document and reduce unfair outputs. |
| GenAI product launch plan | Write a launch plan: users, pricing, risks, monitoring, support and success metrics. |