Anupam Poddar

• Working on LLMS, enhancing and scaling Embedding based Document search method(Retriever Augmented Generation) and fine-tuning with Knowledge graph and key-word search.
• Pdf parsing into chunks using open source pdf parsers.
• POCs around Bio-Bert,Sci-spacy and many other hugging face open source models for bio-medical entity extraction like disease, gene,drug etc.
• Entity ranking with the help Pubmed-Bert,Bio-Bert embeddings for bio-medical context.
• Creation of dataset using techniques like Para-phrasing,Summarization.
• Used Prompt Engineering,Lang chain and open-ai api for answer generation.
• Skills - LLMs, NLP, Fine-tuning, Knowledge Graph,RAG,Chat PDF,Gen AI,Hugging face, BERT,GPT,open-ai,Neo-JS,Vector db,Pinecone,LLama Index, Llama2,Prompt Engineering.Langchain,Pubmed etc.

• Recommender system for load and lorry and network from scratch.

Automating Cash-Back System for Applications (OCR Techniques)

• Utilized OCR techniques and AWS Textract to automate cash-back system, reducing manual EWAY bill verifications by over 60%.
• Successfully extracted RC verified lorry numbers and GST numbers, improving data quality by identifying duplicates and verifying QR codes.
• Designed and implemented a data pipeline to create a dump of EWAY bills between EC2 machines and S3 buckets.

Fraud Detection and Analysis (OCR Techniques)

• Implemented OCR techniques and AWS Recognition to identify fraudulent users by analyzing QR code duplicates and identifying multiple identical EWAY bills.
• Conducted in-depth analysis of more than 100 call recordings to identify patterns of fraudsters and developed a hypothesis for fraud detection.

Call Recording Transcription Analysis (Data Analysis)

• Collected and analyzed around 50,000 call recordings separately for potential fraud users and power users to derive actionable insights.
• Utilized data analysis techniques to identify patterns and behaviors of fraudsters, leading to improved fraud detection and prevention measures.

Real-time Inflow Issue Detection (Data Automation)

• Developed a script to automatically identify inflow issues in important production tables and notify DevOps teams for real-time rectification.
• Successfully scheduled and deployed the script within a day, ensuring smooth data flow and data integrity.

Data Pipelines (AWS, BigQuery, Redshift)

• Created various data pipelines for moving data between BigQuery, Redshift, and S3, enabling seamless data transfer and analysis.
• Designed and implemented a pipeline for collecting call data, improving data accessibility and analysis capabilities.

Aadhar Card Number Masking (Object Detection)

• Leveraged OpenCV and YOLOv8 model architecture to develop an object detection model for masking Aadhar card numbers in images.
• Achieved 100% mean average precision (mAP) for the test set, ensuring data privacy and compliance.

Predictive Modeling and Customer Support Automation

• Utilized BigQuery ML to predict user behavior based on the first 6 and 48 hours of activity, enabling targeted campaigns and better customer engagement.
• Integrated app scripts for customer support, replacing CRM systems for efficient calling and campaign management.
• Extracted and assessed data from databases to drive improvement of product development and business strategies and processes.

• OCR techniques to extract text from warning/error screenshots ongoing applications and applying NLP techniques to remove unwanted symbols and terms and processing it for AUTO heal application to identify error and fix itself.
• Creating rest API with help of flask for predicting diseased olive crops using YOLO model, from labeling and training model to predict it.
• Building an Application to automate Jira test case using techniques custom NER and meaning generation with pos tagging[POC] and development & authentication of complete application using flask API.[OBS]

• Performed sentiment analysis of website texts and built a web scraper using Python, which helped fine-tune market-strategy
• Developed a loan default model using various ML algorithms on a loan book ; achieved best precision scores .
• Performed extensive documentation of Elasticsearch and implemented complex queries and aggregations to form summaries

• Creating End to End Pipeline of current model, Including new Functionalities for product.
• Increased accuracy of Speech recognition model.
• Creating End to End Pipeline of current model, Including new Functionalities for product
• One word Detection Engine (KERAS Model) [PINNA.AI](9 months)

• Performed Web scraping of godown data all over India state-wise, Retail sh ops address scraping from Google maps data, using Google map find landmarks, distance using geographical information and for various purposes.[EDGISTIFY]
• Collecting Godown, warehouse information from newspaper & social media ads using open vision techniques, labeling and categorizing companies/retail shops, Data cleaning using pandas.[EDGISTIFY] (2 months)

Data Science Projects:

• Face-Recognition Attendance System: Leveraged OpenCV and Python to build system that uses facial parameters (e.g., distance between eyelids, cheekbones) to recognize faces and track attendance.
• Skillset Ontology: Developed machine learning model that categorizes projects based on students' technical and non-technical skills, promoting mutual learning. This project incorporated NLP, NLTK, formal concept analysis, and web scraping (Selenium, BeautifulSoup, Mechanize). research paper was published based on this work.
• Fine-tuned Roberta on IMDB dataset for sentiment analysis using PyTorch and achieved best accuracy score of 0.92 • Developed an LSTM sentiment analysis model on IMDB dataset with parts of speech tags and improved accuracy by 4%
• Fine-tuned a Multi-Label Deep Neural Network on top 5 snapshots for relation prediction; developed baseline hits@1 score Garment-search-Engine: Using Feature likelihood method
• Hangman AI - Using LSTM neural net trained model on 250000 words to predict next word.