Pfizer Externship via Extern

Early-career technologist focused on AI for document data

I build AI approaches to extract and structure data from documents, recently through a Pfizer externship in document insights and extraction.

View my work Message Me

Work samples

Portfolio items focus on AI methods applied to documents, including extraction pipelines and insight generation from unstructured text, informed by my Pfizer externship experience.

CINEIQ 6 BEYOND STATIC RECOMMENDATION
CINEIQ 6 BEYOND STATIC RECOMMENDATION
Trained an NCF recommender with BPR ranking loss on 33.7M interactions using AWS SageMaker, achieving NDCG@10 = 0.80, with automated S3 artifact sync. Built a LightGBM demand forecaster (82% RMSE improvement over a naive baseline) and a LinUCB contextual bandit for exploration1exploitation; integrated the Claude API for real-time personalized explanations. Orchestrated the MLOps pipeline: S3 artifact versioning, CloudWatch monitoring and one-command reproducible SageMaker training across 4 components.
PythonLightGBMPyTorchAnthropic Claude APIAWS SageMakerMLflowStreamlit
DISASTERSENSE 6 MULTIMODAL DISASTER SEVERITY DETECTION
DISASTERSENSE 6 MULTIMODAL DISASTER SEVERITY DETECTION
Fine-tuned EfficientNet-B0 (64% accuracy, up from a 58% baseline) and Twitter-RoBERTa (75% accuracy) on CrisisMMD, handling 4.65x class imbalance with weighted loss and differential learning rates; validated per class with precision/recall. Deployed the multimodal severity-scoring system end to end: FastAPI at ~200 ms CPU latency, Docker, a public Hugging Face Spaces demo, and live PostgreSQL-backed monitoring in Metabase.
PythonPyTorchHuggingFaceFastAPIPostgreSQLDocker
SPEECH EMOTION RECOGNITION
SPEECH EMOTION RECOGNITION
Extracted HuBERT embeddings from 7,462 CREMA-D clips and trained a BiLSTM-attention classifier, reaching 71.17% accuracy and 0.72 macro F1 across 6 emotions, 4x the random baseline. Built an audio pipeline from raw waveforms through augmentation (pitch shift, time stretch, Gaussian noise), feature extraction and per-class evaluation.
PythonPyTorchHuBERTBiLSTMtorchaudiolibrosa

About me

I build AI approaches to extract and structure data from documents, recently through a Pfizer externship in document insights and extraction.

I am Asmita Gupta. I have some work experience and I'm exploring how to start my career. I completed the Pfizer Advanced: AI-Powered Document Insights & Data Extraction externship, where I worked on document insights and data extraction using AI methods.

Externships

Pfizer Advanced: AI-Powered Document Insights & Data Extraction Externship

Pfizer

Experience

SOFTWARE ENGINEER (AI/ML) INTERN

AMOTIONS INC. · SEP 2026

SOFTWARE ENGINEER

APPDIRECT PVT LTD · JUL 2022 3 DEC 2023

Education

UNIVERSITY OF MASSACHUSETTS AMHERST

MS BUSINESS ANALYTICS · Class of 2026

PUNE UNIVERSITY

BE COMPUTER SCIENCE · Class of 2022

Skills

AI-powered document insightsData extraction from documentsWorking knowledge from externship projectsPractical experience with AI workflows

Back to work

CINEIQ 6 BEYOND STATIC RECOMMENDATION

Trained an NCF recommender with BPR ranking loss on 33.7M interactions using AWS SageMaker, achieving NDCG@10 = 0.80, with automated S3 artifact sync. Built a LightGBM demand forecaster (82% RMSE improvement over a naive baseline) and a LinUCB contextual bandit for exploration1exploitation; integrated the Claude API for real-time personalized explanations. Orchestrated the MLOps pipeline: S3 artifact versioning, CloudWatch monitoring and one-command reproducible SageMaker training across 4 components.

PythonLightGBMPyTorchAnthropic Claude APIAWS SageMakerMLflowStreamlit

View all work

Back to work

DISASTERSENSE 6 MULTIMODAL DISASTER SEVERITY DETECTION

Fine-tuned EfficientNet-B0 (64% accuracy, up from a 58% baseline) and Twitter-RoBERTa (75% accuracy) on CrisisMMD, handling 4.65x class imbalance with weighted loss and differential learning rates; validated per class with precision/recall. Deployed the multimodal severity-scoring system end to end: FastAPI at ~200 ms CPU latency, Docker, a public Hugging Face Spaces demo, and live PostgreSQL-backed monitoring in Metabase.

PythonPyTorchHuggingFaceFastAPIPostgreSQLDocker

View all work

Back to work

SPEECH EMOTION RECOGNITION

Extracted HuBERT embeddings from 7,462 CREMA-D clips and trained a BiLSTM-attention classifier, reaching 71.17% accuracy and 0.72 macro F1 across 6 emotions, 4x the random baseline. Built an audio pipeline from raw waveforms through augmentation (pitch shift, time stretch, Gaussian noise), feature extraction and per-class evaluation.

PythonPyTorchHuBERTBiLSTMtorchaudiolibrosa

View all work