Prisha S.

Key Skills

Software

Don't disclose

Top Subject Matter

Healthcare / Medical information retrieval (RAG)

Speech / Multilingual language identification

Synthetic data generation for model training / instruction tuning

Top Data Types

Text

Audio

Top Task Types

Prompt + Response Writing (SFT)

Transcription

RLHF

Freelancer Overview

Medical RAG Chatbot project (real-time interaction with iterative feedback-driven refinement). Brings 5+ years of professional experience across complex professional workflows, research, and quality-focused execution. Core strengths include Don't disclose. Education includes Bachelor of Technology, IIT Mandi (2026) and Senior Secondary School Certificate, DPS Civil Lines, Aligarh (2022). AI-training focus includes data types such as Text, Audio, and Medical and labeling workflows including Prompt + Response Writing (SFT), Transcription, and RLHF.

Labeling Experience

HCLTech Hackathon 2025 (Voice of the Nation) — multilingual language identification system

Don't discloseAudioTranscription

Built a real-time language identification system spanning 10 Indian languages using robust multilingual speech classification methods. Designed custom feature extraction algorithms to support accurate classification across multilingual speech data. Validated system performance with high accuracy in multilingual environments, enabling transferable use for multilingual data annotation and language-model evaluation. • Developed and tested feature extraction for multilingual speech signals. • Trained/validated models for language identification across 10 languages. • Benchmarked accuracy across varied multilingual environments. • Produced outputs suitable for downstream multilingual annotation and evaluation workflows.

2025 - 2025

Medical RAG Chatbot project (real-time interaction with iterative feedback-driven refinement)

Don't discloseTextPrompt + Response Writing (SFT)

Implemented LLM-driven retrieval-augmented generation for healthcare queries to produce grounded responses using structured medical corpora. Designed iterative feedback loops for real-time interaction workflows that mirror AI output evaluation cycles. Focused on improving factual grounding and reducing hallucinations through curated input and evaluation-driven refinement. • Retrieved relevant context from structured medical corpora for each query. • Applied prompt-based generation to ensure precise, grounded healthcare answers. • Used iterative refinement based on interaction feedback to improve output quality. • Evaluated responses for factual grounding as part of the RAG workflow.

2025 - 2025

Fine-Tuning LLMs for Synthetic Tabular Data Generation (LoRA, GReaT, I-LoRA; prompt-based sampling)

Don't discloseTextRLHF

Fine-tuned GPT-2 for synthetic tabular data generation using LoRA-based methods and importance-guided neuron selection. Developed I-LoRA to reduce trainable parameters while maintaining strong discriminator performance against GAN and VAE baselines. Applied prompt-based sampling strategies aligned with instruction-tuning and RLHF-adjacent data curation workflows. • Fine-tuned GPT-2 using LoRA and GReaT-style approaches for synthetic data generation. • Implemented importance-guided LoRA (I-LoRA) with gradient-based neuron selection. • Used prompt-based sampling strategies for generating training/generation samples. • Achieved strong discriminator scores to support higher-quality synthetic data for training pipelines.

2023 - 2024

Solar Flare Intensity Classification using ML (benchmarking + SMOTENC oversampling)

Don't discloseText

Built and benchmarked machine learning pipelines to support AI model assessment and dataset quality improvement. Applied SMOTENC oversampling to address class imbalance and improve minority-class performance, directly impacting the quality of labeled datasets used for downstream modeling. Conducted systematic benchmarking across multiple model families to inform evaluation practices. • Benchmarked CatBoost, XGBoost, Random Forest, kNN, and neural networks. • Used SMOTENC oversampling to correct class imbalance. • Improved minority-class performance as part of dataset-quality enhancement. • Produced comparative evaluation outcomes central to model assessment workflows.

2022 - 2024

EEG to Text Converter — iterative benchmarking and signal filtering for improved outputs

Don't disclose

Developed a neural-network prototype to convert EEG brainwave signals into readable text outputs. Applied advanced signal filtering and iterative benchmarking to refine model performance and output usability. Treated the prototype as a training-and-evaluation workflow to improve alignment between noisy input signals and textual predictions. • Performed signal preprocessing and filtering on EEG inputs. • Iteratively benchmarked model outputs to guide improvements. • Refined model performance through repeated evaluation cycles. • Produced text outputs suitable for assessment and potential dataset refinement.

2022 - 2023

Education

Work History