Shraddha
Suresh Kadam

Individual Assignment · UCD

• Designed a 6-experiment controlled study comparing Direct, Chain-of-Thought, Role-Based, and Combined prompting strategies across three document genres.
• Controlled temperature (0.1–0.7), system role, and document type — isolating each variable's effect on sentiment output quality.
• Key finding: CoT + Role-Based strategies significantly outperform direct prompting for detecting sarcasm and implicit emotion.
• Documented failure modes: high-temperature hallucination, over-confident labelling, strategy-content mismatch.

Group Project · UCD

• Led full RAG system setup on Stack AI — Claude 4.5 Sonnet, semantic chunking, sliding window memory (size=10), citations enabled.
• Engineered 5-rule system prompt for safety: document-only grounding, anti-hallucination, out-of-scope refusal, no medical advice, citation enforcement.
• Designed 24-test evaluation framework (Factual / Synthesis / Edge-Case, 1–5 rubric, multi-rater scoring) — 4.3/5 factual, 3.9/5 synthesis & edge-case.
• Identified confident hallucination on edge-cases as primary failure; produced production improvement roadmap.

TechSaksham · Microsoft & SAP / Edunet Foundation

• Built NLP chatbot in Python — end-to-end from intent design to model integration.

Bhabha Atomic Research Centre (BARC)

• SSL-VPN log analysis system in PHP — security-critical research environment.

TagSocial, Syneidisi Tech

• Visual communication and brand design — translating complex ideas for diverse audiences.

University College Dublin

• LLM prompting, RAG systems, NLP evaluation, HCI research, AI systems.

Terna Engineering College

• Algorithms, ML fundamentals, software engineering · NSS Secretary, 100+ volunteers.