Skip to content
View yamini-nlp's full-sized avatar

Block or report yamini-nlp

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
yamini-nlp/README.md

Yamini G

Google Scholar ORCID LinkedIn Portfolio

CS graduate. I study how AI systems fail when sentiment, confidence, and surface fluency masks the signal that actually matters — user intent, factual grounding, and decision-relevant uncertainty.

Research

Three preprints on sentiment analysis, intent modeling, and human-AI communication. All self-initiated during undergraduate study.

Paper Venue Links Implementation
Sentiment-Aware Reflective Writing Systems — formalizes the gap between sentiment polarity S(x) and intent interpretation I(x,C,G,H); proposes utility-theoretic response selection with configurable asymmetric cost weights TechRxiv · IEEE DOI MindNook (pinned)
Beyond Surface Affect — proves formally that sentiment detection does not imply intent interpretation; identifies four canonical failure modes with deployed prototype observations TechRxiv · IEEE DOI Under active development
YouTube Transcript vs Comment Sentiment — dual-model pipeline documenting divergence patterns across five content domains; failure modes in public discourse analysis SSRN DOI Repo

Projects

Project What it addresses
LLM Reliability Lab Hallucination benchmarking across 3 live LLMs via Groq (Llama 3.1/3.3, GPT-OSS 120B) — dual heuristic + LLM-as-judge scoring, Wilson confidence intervals. CoT reached 87.5% vs. 85.0% zero-shot accuracy (n=40, overlapping CIs).
MindNook Prototype of published TechRxiv framework — five-layer NLP architecture with utility-theoretic action selection and ethical filter
Fake News Detector Multi-signal fusion: XLM-RoBERTa + Google Fact Check API + propagation graph features on LIAR dataset
PrognosAI Clinical NLP: 30-day readmission, LOS, and specialty prediction from discharge notes — three pipelines (TF-IDF, vitals-hybrid, LLM), SHAP/phrase-level explainability, CI + Docker
Prism RAG pipeline with FAISS vector search and claim-level hallucination detection — unsupported claims flagged before reaching the user; Dockerized backend/frontend with a real evaluation harness (Recall@5, groundedness rate, Wilson CIs)
PrepSphere AI placement-prep platform with dual LLM reliability paths — live Groq proxy + BullMQ async queue with Zod schema validation. Node.js · MongoDB · Vanilla JS.

Technical Areas

NLP Transformer Fine-tuning LLM Evaluation Hallucination Detection Retrieval-Augmented Generation Clinical AI Sentiment Analysis Intent Modeling PyTorch scikit-learn FAISS FastAPI Next.js

CS graduate · 3 preprints · 6 deployed systems · NLP & LLM reliability

Pinned Loading

  1. MindNook-HCJ MindNook-HCJ Public

    Prototype implementation of published TechRxiv framework — five-layer NLP architecture with sentiment detection, intent modeling, and utility-theoretic action selection

    HTML

  2. llm-reliability-lab llm-reliability-lab Public

    Medical-QA hallucination benchmark for 3 live LLMs via Groq (Llama 3.1 8B, Llama 3.3 70B, GPT-OSS 120B) — dual keyword + LLM-as-judge scoring, Wilson confidence intervals, phrasing-ambiguity axis. …

    TypeScript

  3. PrepSphere PrepSphere Public

    AI placement-prep platform built as a controlled comparison of two LLM reliability strategies: a live Groq proxy vs. a BullMQ async queue with Zod schema validation.

    HTML

  4. Prognos-AI Prognos-AI Public

    Clinical NLP: 30-day readmission, length-of-stay, and specialty prediction from discharge notes using three independent pipelines — TF-IDF baseline, vitals-hybrid, and [exact Groq model string] wit…

    Python

  5. fakenews_detector fakenews_detector Public

    Multi-signal misinformation detection: XLM-RoBERTa fine-tuning + Google Fact Check API + propagation graph features, fused via weighted ensemble on LIAR dataset

    Python

  6. prism prism Public

    Retrieval-Augmented Generation platform with FAISS-based semantic search, claim-level hallucination detection, and confidence scoring. Includes a real evaluation harness (Recall@5, groundedness rat…

    TypeScript