Anshuman Guha Email and Phone Number

(Released as Freddy Copilot feature)● Proactive Quality Coach: Led Gen-AI-based real-time write-assist product for agents, improving grammar, spell, relevance, tone, abuse, filler words and length. Deployed LoRA fine-tuned Llama-3-8B & other small LMs with inference performance optimizations (quantization & flash attention). Presented at DSS-SF and AIAI-Austin & received press-coverage from the company's editorial columnist.● Post-Resolution Quality Coach Spearhead feature development to score agents’ performance on agent-user interaction data. Instruction fine-tuned 7B Mistral LLMs on training data generated by GPT-4 and human annotations(Research led at AI Labs) ● Instruction Tuning of LLMs : Extended Grammarly CoEDIT research paper to seven languages with Llama-8B, productizing the Proactive Quality Coach for text enhancements. Fine-tuned using LoRA for diverse instructions, like "Improve with minimum required edits preserving native ascents". Implemented custom huggingface logit processors to prevent generation of nuisance edits.● Unified Text-Search on FAQs: Developed an advanced RAG bot for FreshChat/Freshdesk enterprise FAQs, improving search result precision by 30% and relevance by 25%. Using LlamaIndex finetuning infrastructure Fine-tuned embeddings and generator, enhanced query pre-processing, hybrid search, and advanced chunking. ● Open sourced: Token Efficient Indexing with Entity Extraction on conversational data: Optimized vector search with token-efficient indexing and entity extraction. Developed a process involving meaningful data classification and extractive summarization for top-k sentence selection. Implemented a custom NER tagging model to identify food entities and enhanced search with HNSW index tags.

* Spearhead multiple credit-card fraud models deployments with models having annual traffic of up to 20 million users in real-time traffic. Reduced turnaround time for new deployments by 80% with timely releases.
* Render usage of acquired 2-D Tradeline data per customer by building CNN LSTM stacked models for fraud prediction * Contributed to the development of credit-risk and response models. Lead AUROC improving strategies using advanced feature engineering by implementing customer segmentation techniques using Autoencoders and clustering.
* Reduce model building lifecycle by improving training pipeline to enhance re-use, robustness and explainability. This pipeline is used to build credit risk models to predict fraud probability using GBM Models.
* Using ML model scoring deployment pipeline and reduced 99% of ML errors in production using API based docker tests and simulated-data tests. Reduce scoring SLA by 90% to enable acceptance of large ensemble models by developing Apache arrow like columnar Dataframe and used NumPy broadcasting & binary searches.
* Training more than 20 Data Scientists and Senior Managers by curating hand-on training & workshops for productionizing models at scale on real time and batch scoring platforms.
* Co-Authored an article in KDNuggets on “MLOps Best Practices ” with senior management.

Developed and Deployed• Core and leading data scientist to conceptualize, design and normal behavior modelling product for the company. Developed various neural network models like Multi-Channel CNN, sliding window autoencoders etc., to model new normal, ambient conditions, equipment aging, new normal and address scarcity of failure data. Novel feature importance methods implemented to work with time-window based hotelling t-squared error score• Extended work on applied research for transfer learning Beyond Sharing Weights for Deep Domain Adaptation to time-series data. This work helped in exploring approaches where labeled data is not readily available.• Implemented time series forecasting models using statistical feature engineering and deep learning models to predict energy prices for real a day ahead spot market to calculate optimum split between real time trading and day-ahead guaranteed commitment. • Applied density based and hierarchical clustering methods for anomaly detection on time-series data to compensate lack of failure data• Implemented semantic search for work orders on maritime companies work-order system to create technical knowledge base for future• Worked with product & customer success leadership, clients and software engineering teams to prioritized resources for MVP’s feature requests with internal product development and feature requests from sales and customer support.Articles and Publications• Presented my research at OSDC West Conference (2017) focusing deep learning modeling techniques for failure prediction (Video Link)• Co-authored blog for company’s white papers for successful proof-of-concept ML model development use-cases

Responsible for implementing data mining and statistical machine learning solutions to various asset management problems such as for predicting equipment failures, remote monitoring of current health, energy consumption and life-cycle maintenance costs:• Implemented multivariate regression models for equipment fault detection system. Used Lasso feature selection with multicollinearity tests. Achieved more than 40% improvement in RMSE over legacy models. Model deployment allowed safety of power generation system for ships • Execute bi-weekly post-deployment model monitoring tasks and prepared model performance reports and recommended action items for operations staff• Developed diesel engine performance index using mahalanobis distance measure of engine parameters from healthy & new engine• Led initiative to create statistical models using historical data to predict ship’s engine exhaust temperatures in several global climatic conditions• Developed faulty injector detection models relying on classification algorithms using cylinder pressure, temperature and other parameters from online monitoring tools such as diesel doctor• Applied data mining to maintenance task deferral problem which demonstrated potential savings of maintenance dollars across fleet• Prototyped a predictive model using real-time turbocharger rpm data, to optimally schedule periodic cleaning