Developed and optimized enterprise-grade prompt engineering frameworks using LangChain and LangGraph, resulting in a 40% improvement in content generation quality for sales and marketing materials across multiple industry verticals.

Led knowledge management initiatives including fine-tuning Llama models on domain-specific data, building structured knowledge graphs, and designing information retrieval systems.

Collaborated with cross-functional teams to align prompt engineering strategies with business goals and improve overall content performance.

Developed hallucination detection and content verification systems using LangChain agents and LangGraph, creating ensemble techniques that reduced inaccuracies by 65% while maintaining strict data privacy protocols for enterprise clients.

Engineered prompt frameworks for claim extraction and verification using OpenAl Assistants and custom Python tooling, enabling automated quality assurance processes that scaled to process 30,000+ content pieces daily.

Implemented and fine-tuned Llama models as verification agents, developing novel techniques for cost-effective factuality assessment that maintained 92% accuracy while reducing API costs by 72%.

Extracted key medical claims from research documents and materials published by U.S. pharmaceutical companies, meeting FDA requirements.

Used extracted data to create presentations for customers and clinicians, supporting sales pitches for new drugs.

Built the solution on LangChain, with fine-tuned models and prompt engineering.

Developed Al-powered social media app features for personalized content creation using Stable Diffusion and Deepgram.

Enabled near real-time immersive experiences with professional model photos, videos, and voice cloning.

Used machine learning techniques like low-rank adaptation (LoRA) based on PEFT, model quantization, checkpoint merging, and other related approaches to achieve state-of-the-art (SOTA) optimization and performance.

Completed a POC of a mental health assistant built on top of Voiceflow. It understands the user's mind, responds with appropriate custom suggestions from the given knowledge base, or engages in a general conversation like a human.

Built the RAG-based chatbot that replicates the behavior of a neuroscientist and emphatically engages with the user during the conversation.

Used LangChain to enrich the vector store, adding proprietary neuroscience and psychological data to generate a more relevant response from the LLMs.

Built a recommendation engine for an addiction recovery app that incorporates user demographics, mental state, and other metadata from app interaction to suggest quotes and prayers relevant to the user's state of mind.

Added memory and linked the vector database quickly.

Provided several offline and online metrics to track the live performance of the model and the recommendation system.

Trained causal language model (CLM) with DeepSpeed on Bittensor, a peer-to peer open-source protocol powering a globally distributed decentralized neural network.

Used a mountain dataset (800+ GB) to train generative LLMs using DeepSpeed, which has lower network validation loss and thereby increases the model's reward, benefiting the client.

Contributed to hyper-parameter tuning and revamped the default PyTorch data loader for faster dataset loading. Also, validated the models' performance on the validation set to ensure it meets the desired performance metrics.

Developed a user sticker model using Stable Diffusion and ControlNet for a given style.

Experimented with different configurations of parameters to understand the effect on the results and achieved state-of-the-art results.

Explored different methods like checkpoint merging, trained a stickers model, and then fine-tuned that model on Portrait Face, and combined sticker style and user image training.

Revamped roof measurement automation for US houses by engineering a high-performance Line Detection model on aerial imagery to detect roof, wall, and ground lines from 28% to 59%.

Developed segmentation models for roof and wall facet detection while handling obstruction, achieving 86% accuracy.

Orchestrated microservices using Step Functions, AWS Lambda, and Amazon SageMaker endpoints, improving response time by 25%. This was followed by simplifying maintenance and support, focusing on maximizing code reuse.

Designed and built a chatbot using Hugging Face transformers with DeepSpeed to resolve queries for over 400 million of active monthly users; deployed successfully, handling over 10 million daily conversations with 85% accuracy.

Improved the accuracy of face recognition, matching, and liveness-check models, added face segmentation, blur and forgery detection, real-time blacklisting, and face search, enhanced service efficiency, and reduced manual effort by over 25%.

Inculcated best MLOps practices on the team to lower technical debt and adapt workflow that requires little to no manual intervention for simulations, testing, deployment, and monitoring.

Designed and deployed an OCR solution that outperformed the accuracy of Google OCR by 12% as benchmarked on the ICDAR 2013 dataset.

Prototyped an NLP--based bot using recurrent neural networks to parse comments on the website, assess sentiments, and classify them into themes.

Worked on a trained convolutional neural networks model using INT8 fixed-point and trained a software categorization and brand recognition tool using word2vec, N-grams embeddings, and kNN classifier.

Reduced the overall cost from an estimated $1 million to less than $50,000 per year.

Enabled standard responses and activation of the needed systems to respond.

Conceptualized and built a ConvLSTM model on TensorFlow for hand gesture recognition.

Worked on single-shot face recognition using SSD MobileNet for detection and Inception-ResNet v2 for feature extraction.

Led a team to propose and implement predictive analysis for retail sales forecasting using ARIMA, exponential smoothing, and Holt's Winter models, resulting in a 5.35% increase in quarterly sales.

Used Bayesian optimization techniques for hyperparameter tuning.

Introduced TP-GAN for photorealism and identity, preserving the frontal face-view synthesis from any pose.