Pranav

Bristol Myers Squibb AI-Hackathon Dynamic ML Pipeline for mAb Fucosylation Prediction

• Built a full-stack ML dashboard predicting mAb fucosylation from 10 bioprocess variables, training 7 models (Ridge, PLSR, Random Forest, XGBoost, GPR, ANN, Hybrid), achieving R²=0.83 with XGBoost+SHAP on 10,000 samples and demonstrating 1,440% ANN recovery by scaling from 500 samples, proving neural network incompatibility with small pharma datasets; integrated Gemini 2.5 Flash for ICH Q8/Q9-aligned regulatory report generation.
• Engineered passwordless authentication (Google OAuth 2.0, Gmail SMTP OTP with signed token sessions) and a full support ticketing system as part of a production-grade React + FastAPI deployment.

Reference Link

Bristol Myers Squibb AI-Hackathon Offline Computation Tool to Estimate Charge Variants of a Biopharmaceutical Protein - 2025

• Estimated charge variant distributions using an offline Python tool with strict data validation, applying convolution, polynomial powering, and FFT-based convolution, with moment-matching approximation for scalability, and stresstested on over 100 profiles, evaluating 10–50 charge states per run with stable runtime and memory.
• Generated PMF plots, charge state distributions, and summary statistics, enforcing no renormalization to preserve 100% probability mass within the truncation window.

Reference Link

Glycan Clearance Analysis & Visualization for Therapeutic Optimization - 2025

• Designed a dynamic Power BI dashboard with 6+ analytical visualizations (stacked bar plots, line charts, card visuals) to analyze glycan clearance trends across 5 production sites (A–E), enhancing clearance prediction accuracy by 20%.
• Applied K-means clustering (k=3, Elbow Method) on 50+ glycan profiles, classifying them based on AUC and half-life.

Reference Link

Digital Student Attendance Tracking System - 2025

• Developed an Attendance Tracking System using PostgreSQL and a Python/Flask web app, implementing foreign key constraints, hashed authentication, and role-based portals for Admin, Instructor, and Student across 7+ relational tables.
• Built a web-based UI supporting full CRUD operations and session-level attendance tracking with 3 statuses (Present/ Absent/Late), implemented cascaded deletion to remove dependent records, and added CSV import/export + downloadable attendance summaries for quick reporting.

Reference Link

Conference based Agenda Optimization: Identifying Similar Abstracts - 2025

• Detected duplicate/highly similar conference abstracts (2014–2025) using text tokenization, TF-IDF & cosine similarity (>0.85) on 254K+ entries (30 features), leveraging Depth First Search graph traversal for node transitivity to cluster related abstracts into connected components, including isolated entries.
• Built a Python pipeline and Tableau dashboard for automated preprocessing, similarity scoring, and multi-year clustering to optimize agenda planning.

Reference Link

HLS based Signal Modulation Classification System using Deep Learning - 2024

• Used CGRDNN, a hybrid neural network for recognition of signal modulation through high-level synthesis, facilitating practical FPGA implementation using high-level synthesis and demonstrating ASIC development potential.
• Reduced processing time by 20% compared to conventional models for complex signal classification.
• Trained the neural network to attain 98% accuracy in classifying modulated signals, including those corrupted by random noise, in challenging non-cooperative channel scenarios.

Reference Link