Evaluation, safety & observability
Apply guardrails, safety and privacy controls
Input/output filtering, PII redaction, prompt-injection defenses, responsible-AI policies (DPDP-aware).
~8 focused hours·intermediate
Tools: Guardrails AI, NVIDIA NeMo Guardrails, OWASP LLM Top 10, PII redaction (Presidio-style), DPDP-aware policies
Market relevance — share of job ads asking for this
What employers mean
You should be able to…
- Filter and validate user inputs to block prompt injection and jailbreak attempts before they reach the model
- Redact or block PII (Aadhaar, PAN, phone numbers, bank details) in both inputs and outputs, DPDP-aware
- Validate model outputs against a schema or policy before they're shown to a user or executed as an action
- Detect and block system-prompt leakage/exfiltration attempts
- Apply topic/scope restrictions so a domain-specific bot refuses out-of-scope or unsafe requests gracefully
- Red-team your own LLM app for prompt injection, data exfiltration, and unsafe tool invocation before shipping
- Write and enforce a responsible-AI policy covering what the system will and won't do, with a human-escalation path
- Layer guardrails so a failure at one layer (e.g. model refuses) still fails safely rather than crashing or leaking
Needs first: Integrate LLM APIs into an application
Learn — free, link-checked
The few resources that matter
Read · intermediate · 15 min · docs.anthropic.com
Reduce prompt leak
Concrete techniques to stop system-prompt exfiltration, one of the most common guardrail asks in Indian AI interviews. — Anthropic
Read · intermediate · 30 min · guardrailsai.com
Introduction - Guardrails AI
Open-source library for input/output validators (PII, toxicity, schema) you can drop into any LLM pipeline. — Guardrails AI
Read · intermediate · 40 min · docs.nvidia.com
NVIDIA NeMo Guardrails Library Developer Guide
Programmable rails (topical, safety, jailbreak) with Colang, a second widely-cited guardrails framework alongside Guardrails AI. — NVIDIA
Read · intermediate · 45 min · owasp.org
OWASP Top 10 for Large Language Model Applications
Industry-standard checklist (prompt injection, insecure output handling, excessive agency) interviewers expect you to know by name. — OWASP Foundation
Course · intermediate · 90 min · deeplearning.ai
Red Teaming LLM Applications
Teaches how to actually attack your own app (prompt injection, jailbreaks, PII leakage) before shipping guardrails against it. — DeepLearning.AI
Practice
Guardrailed GenAI assistant for an NBFC customer support widget
Build a customer-support assistant for an NBFC that answers loan/EMI questions, with input guardrails blocking prompt injection and PII (Aadhaar/PAN/phone) leakage, output guardrails blocking specific investment advice and system-prompt leakage, and a documented escalation path to a human agent when guardrails trigger. Red-team it yourself with at least 10 adversarial prompts and document what got through before your fix.
Done when
- Input guardrail blocks or redacts PII (Aadhaar/PAN/phone patterns) in at least 90% of a test set of 20 crafted inputs
- Output guardrail blocks specific investment-advice language and system-prompt leakage attempts
- A documented red-team run of 10+ adversarial prompts with before/after results after hardening
- A clear, logged escalation path (message + flag) when a guardrail triggers, instead of a silent failure or crash
Prove it
Evidence a recruiter can check
- Public GitHub repo with guardrail configs/code and the red-team prompt set with results
- A before/after table showing which adversarial prompts were blocked after hardening
- A short written responsible-AI policy doc for the assistant (scope, refusals, escalation)
Interview
Questions you'll get asked
- How do you defend an LLM app against prompt injection from untrusted retrieved documents or tool outputs?
- How would you redact Aadhaar/PAN numbers from both user input and model output in a DPDP-compliant way?
- Walk through the OWASP Top 10 for LLM apps — which 3 have you actually had to defend against?
- How do you stop a customer support bot from being tricked into revealing its system prompt?
- Design guardrails for a financial-advice chatbot that must never give specific investment recommendations.
- What's your process for red-teaming an LLM app before launch?
- How do you balance guardrail strictness against false-refusing legitimate user requests?