As a Senior Machine Learning Engineer working on Document AI I leveraged my expertise in computer vision, optical character recognition (OCR), and natural language processing (NLP) to develop solutions for intelligent document processing (IDP).- Designed and implemented robust key-value pair extraction models for structured and semi-structured documents, enabling efficient data extraction and processing.- Evaluated and integrated state-of-the-art OCR engines, such as Google Vision API, Amazon Textract, PaddlePaddle, and EasyOCR, to ensure accurate text recognition from various document formats.- Collaborate with cross-functional teams to understand client requirements and deliver tailored solutions for clients in UAE and Southeast Asia.- Continuously researched and implemented advanced techniques in machine learning, deep learning, and computer vision to enhance model performance and accuracy.- Optimize models for scalability, performance, and deployment in production environments.- Contribute to the development of reusable pipelines, and frameworks for document AI tasks.- Mentor and guide junior team members, fostering a culture of knowledge sharing and continuous learning.-Actively worked on POCs involving prompt engineering, Retrieval Augmented Generation( RAG)using vector embedding databases and fine-tuning LLMs using PEFT.-Simultaneously leading the development of No Code ML & DL Platform ( Image, Text, Tabular, Documents ) leveraging stateless architecture and all the open source tools for experiment tracking, project and experiment management, model deployment, serving, and monitoring integrated into the platform itself.- Managing a team of ML engineers and analysts and working with UI developer, UX designer, and backend developer to build and scale ML platform

-Spearheaded development, deployment, and automation of India dashboard, consumed by CEO office, Marketing, Risk, and CS verticals. -Designed and developed an Analytics Layer that is highly critical for repayment reconciliation of $1.3 Million and dynamic DPD calculation for 10k+ loan accounts and thus increasing the monthly collections by 23%.-Designed a sandbox for data sanity supported by data of different time periods around several business metrics definitions and thus improving communication within the whole enterprise.

Led a team of 10 associates. Was responsible for everything related to data of a data enrichment product named Algo360 with annual revenue INR 9 cr.- Built a custom text processing pipeline for SMS data, that handled a data volume of 200 GB/daily, resulting in improvement in Weighted-F1 scores for 7 out of 16 classification models. - Standardized and sacrosanct the POC process for new client onboarding and eventually removed all manual interventions. - Fine-tuned spacy's pre-trained NER model for entity extraction from SMS text.- Experimented with the following model and text representation combination for model re-training and improvement :1. TF-IDF and (Logistic Regression, Decision Tree, Random Forest, XGboost)2. Word2Vec & Mittens with LSTM with different layer combinations. - Created data cleaning and annotation process through the integration of Open Refine and use of cologne-phonetic, Metaphone, and fingerprint clustering algorithms, for training data and thus reducing the manual effort of data annotators by 250X. - Developed LSTM-based deep learning models as replicas for all 16 models in production and run parallel experiments to measure latency when deployed using a docker container. - Monthly monitoring and validation of retail lending scorecard developed using alternate data produced by algo360.- Developed semi-supervised (combination of rule and clustering algorithm) algorithm for exception handling mechanism to identify a shift in data distribution.- Managed, guided, and mentored associates for ad-hoc analysis and feature enhancements. - Performed POC for Algo360's sister product based on E-mail data.- Performed POC for Account Aggregator (Setu) as another data enrichment source so as to improve the accuracy of the data points.

• Closely worked with client SVP and CIO.• Developed a DBSCAN model to identify new shopping hot spots in 11 cities based on credit card transaction data and thus improving the credit card offer rollout in physical stores. • Built a custom algorithm for tagging the geo-location of merchants' physical stores. The algorithm was built using the triangulation method and proxy for transaction graphs on one-month transaction data across 11 states of India. The algorithm was capable of incremental tagging in each iteration with new transaction data. The algorithm was built to facilitate real-time offer recommendations to customers when they swipe at any location. The then price of such a product in the market was around INR 5cr. Later on RBI came up with frame work for geo-tagging in Oct 2021. RBI Notification Link https://www.rbi.org.in/Scripts/NotificationUser.aspx?Id=12260&Mode=0 • Developed an xgboost model to reduce the load on the dedupe underwriting queue with less than human error rate (1%) and thus reducing 39% the stress of the queue. The model was used for the automatic segregation of unique applications only and did not interfere with non-unique/duplicate applications. It was a model for the operations team and not the risk team. • Performed a POC to predict the delinquency bucket switch of a customer and thus improved upon the previous model by reducing the f1-score rate by 50%. After delivery of POC; SOW for the project worth $ 1.9 Million was signed.

• Developed a Voice-enabled Information Retrieval System using weighted results of the analytical module of Apache Solr and Manhattan LSTM model – a Siamese network trained on FAQ documents, to assist customer support executives, for 3 LOBs of client.- Speech-to-text was performed using Google's web speech API. - FAQ documents were dumped in MongoDB as the primary Database and indexed in Solr for search operation. - Built a connector between Solr and MongoDB for automatically fetching and indexing documents uploaded in MongoDB. - Used Quora-Question Pair Similarity data for training Siamese Network and then used transfer learning for fine-tuning embedding layers with FAQ document. - Manually tuned analytical module of Solr using various text-processing components in the pipeline.• Performed grouping for car parts using several unsupervised clustering techniques ( fingerprint, metaphone etc. )to reduce the total unique parts count from 150k to 80 groups on the basis of name similarity, cost range and position of car parts in the car.

Designed & implemented modules for data sharing with Credit Reference Agencies (Equifax, Experian) to reduce barriers to entryand lender safety for small-scale lenders by providing credit score.• Designed & implemented an automatic data flow pipeline for data migration from different sources into Data Lake, enabling 360view of customer data.