How the AI terms fit inside each other
Each circle is a smaller part of the one around it. Every LLM is generative AI, all generative AI is deep learning, and so on. XAI is not a circle: it is a set of methods for explaining any of them.
The AI vocabulary
Search any term. Each card has a plain definition and an example from our own work.
No matching term. Try a shorter word.
Three ways to make an AI "know" our material
| Prompting | RAG | Fine-tuning | |
|---|---|---|---|
| What it is | Writing clear instructions and examples in the question | Searching our documents first, then letting the model answer from what it found | Further training the model on our examples so it changes its behaviour |
| Best for | Quick tasks, formatting, tone | Facts that must be correct and cited: guidelines, syllabi, SOPs | Style and skill: writing MCQs our way, a consistent tutor voice |
| Cost | None | Low: a vector database and some setup | Medium: GPU time and a curated training set |
| Updating | Edit the prompt | Add or replace a document; instant | Retrain |
| Our example | Vritta's minutes format | Diabot answering only from ICMR/RSSDI guidelines | Planned: a public-health MCQ generator for Pariksha |
Rule of thumb: use RAG for knowledge, fine-tuning for behaviour, and prompting always.
How a web app actually runs
Every tool on our Innovations page follows this path. When something breaks, it breaks at one of these boxes.
The infrastructure vocabulary
No matching term. Try a shorter word.
Local server, cloud server or free hosting?
| Option | Examples | Good for | Watch out for |
|---|---|---|---|
| Static hosting | Firebase Hosting, GitHub Pages | Sites with no server code: this website, PRISM's pages | Cannot run Python or keep a server-side database by itself |
| Platform-as-a-service (PaaS) | Railway, Render | Apps with a backend: Pariksha, Sangam, Artha | Free tiers sleep (slow first load) and may have no backups |
| Backend-as-a-service | Firebase Firestore, Supabase | Login, database and storage without running your own server | Usage-based bills; lock-in to one provider |
| Cloud virtual machine | AWS, Azure, Google Cloud, Indian clouds | Full control, big workloads, GPUs by the hour | You manage security and updates; costs grow quietly |
| Local (on-premise) server | The lab's DGX Spark and Mac mini servers | Sensitive data, private AI models, no monthly GPU rent | Needs power backup, IT network access, and your own backups |
Ten things every digital health builder should know
1. Back up everything, automatically
Follow the 3-2-1 rule: 3 copies, on 2 kinds of storage, 1 off-site. Free hosting tiers often keep no backup, so one wrong "reset" can erase exam results permanently.
2. Never put secrets in code
API keys and passwords go in environment variables or a secrets manager, never in GitHub. A leaked key can run up bills or expose data within hours.
3. Identifiable data needs a reason and a lock
India's DPDP Act 2023 applies to student and patient data. Collect the minimum, get consent, restrict access, and keep sensitive data off public AI services.
4. Own your accounts institutionally
Domains, app-store accounts and hosting should be under the institution, with two admins, so a tool never depends on one person's login.
5. Separate test from live
Try changes on a copy (a "staging" environment) before touching the live exam or portal.
6. Keep code in Git
Git records every change so you can undo mistakes. GitHub stores it online and lets others help.
7. Memory decides which AI model you can run
A model roughly needs (parameters × bytes per parameter) of memory. A 70B model at 4-bit needs about 40 GB before it answers a single question.
8. Validate before you trust
An AI tool is a health intervention. Test it against experts (as Diabot does) and report accuracy before real use.
9. Watch the free-tier limits
Free plans sleep, cap storage and change terms. Fine for demos; use a paid or institutional plan for anything students depend on.
10. A human approves every AI output
Minutes, MCQs, clinical answers: AI drafts, a qualified person checks and signs off.