Suresh Palla

Generative AI insights:

• Fine-tuned and quantized LLM models using QLORA and open-source LORA adapters to generate insights from campaign monitoring dashboards.
• Enhanced model performance through domain-specific data augmentation and Direct Preference Optimization (DPO), improving insight relevance by ~20% based on human evaluation metrics.

Contextual Relation Extraction:

• Performed Supervised fine-tuning of open-source LLMs (e.g., Flan-T5) for relation and dependency extraction from product catalogs, enabling assisted generation of rich and informative ad creatives using Hugging Face, GPTQ.
• Leveraged PEFT (Parameter-Efficient Fine-Tuning) techniques to reduce compute costs by ~30%, enabling low-latency inference.

User Profiling and Segmentation:

• Developed solution to capture customer traits for segmentation and content/ad personalization.
• Employed Latent Dirichlet Allocation (LDA), PCA, Hierarchical clustering techniques to form highly tailored segments which increased average CTR by 150%.

Multi Variate Attribution:

• Developed model in Python and Py Spark based on Random Forest and interpretable ML model (SHAP) to identify key levers for increasing conversions. This helped in improving conversion rates of two major campaigns from 0.3 and 0.6 to 0.7 and 0.9 respectively.

NLP for TV OS:

• Formulated problem of information extraction from TV commands as Intent Classification and Named Entity extraction and used conversational AI frameworks like RASA employing Transformers based DIET classifier, Large Language Models (LLMs) like BERT and Conditional Random Field (CRF) models.
• Improved precision and recall by 65% and 33% respectively compared to baseline model which used primitive NER approach.

Similar Show Extraction:

• Created wider audience segments for shows by coming up with larger set of similar shows using based Sentence Transformer based contextual embeddings and was able to achieve an average of 5.5 times expansion over 15 segments

Personalization Team:

• Improved existing CPU based Item Recommender System using GPU based KNN search approach implemented using FAISS architecture.
• Attained 8x increase in speed and 0.01 increase in nDCG score compared to baseline model.

• Performed feature selection based on Statistical hypothesis testing and LASSO approach and a multi-label classifier based on Naïve Bayes to classify vendor risk bands.
• Developed Sentiment Analysis model using NLTK and Deep learning-based embeddings for analyzing COVID Impact assessment on sales.

• Automation of reporting and analytics dashboards for Credit Risk Assessment and fraud detection.
• Developed Py Spark based routines for data pre-processing and performed exploratory data analysis.

• Led team of 8 members for automating tests for validating exhaustive scenarios in banking and finance domain.
• Achieved 20x faster test execution with test coverage of 80%.