Shridhar Agarwal

Worked as Lead AI Engineer for Online Course System of Cambridge University

• Built a Student Companion System using RAG with capabilities to answer questions from courses (with multi-modal content), verify generated response (using agentic chain), semantic search course content (using vectordb), generate quiz and opt from pedagogical styles
• Built a Automatic Marking System using few-shot prompting which scores short answers (~85% acc) & provides reasoning using CoT

I am the proprietary owner of this firm

• For a travel business, built a chatbot which takes personality quiz and recommends places and activities most suited to their taste
• For a cyber-security client, built an web-extension capable of parsing cyber-security related entities from new-articles
• For a marketing client, built 15+ projects related to recommendation, retention, campaign targeting and customer segmentation
• For a grocery business, built a chatbot which can fetch details about the products and availability from db and answer relevant questions
• Built a sports business, built a Gym-Drill Analysis System recorded drill timing and accuracy using CV models like Mediapipe & YOLOv4

Operated as the sole ML engineer. Built end-to-end ML products.

• Built an Entity Extraction Model using SPACY(~0.78 F1) for extraction of 12 cybersecurity entities. Used combination of NER, sentence classification and rule-based matching. Experimented w. BERT & STANZA models
• Built a Tactic & Technique multi-label classification model using a custom ML pipeline of several LR models being used in sequence & parallel
• Built the application using Python and Django to support ML models. Used multi-tread parallel processing to reduce inference time. Also implemented feedback mechanism for future retraining
• Developed POC of a Text-to-NoSQL system for Elastic Search db using pretrained PICARD model modified to support ES db

• Built Order Forecasting for capacity planning using FB-Prophet
• Built ML Management System w. Airflow , MLFlow for Continuous Training , Registry & Tracking
• Built Table Parser (~80% acc) for PO. Used FasterRCNN for table & row detection & BiLSTM to map column to headers
• Built Prod Desc Matching w. BERT using Triplet Loss.

• Built Transactional Anomaly Detection System on SMACK using rule based statistical models to return risk-score under 300ms
• Built Credit Risk Model (~0.81 F1) with Random Forest on Spark.

• Upgraded db entry system (Stored Procedures SQL) from single to double entry (credit & debit) across 18 banking clients