Soham

Formulated a search content refresh scoring pipeline across 30,000 performance records (32 client sites), stripping temporal trend fields to enforce strict data leakage controls. Evaluated classification architectures under a grouped client holdout split (8 unseen clients, 7,115 test rows) to benchmark model generalization against a heuristic baseline. Boosted review queue precision to 0.80 P@50 (vs. 0.52 baseline) and 0.85 P@20 across held-out client domains.

Machine Learning Engineering Intern at FlyRank AI (2026-07 – 2026-08)

• Formulated a search content refresh scoring pipeline across 30,000 performance records (32 client sites), stripping temporal trend fields to enforce strict data leakage controls.
• Evaluated classification architectures under a grouped client holdout split (8 unseen clients, 7,115 test rows) to benchmark model generalization against a heuristic baseline.
• Boosted review queue precision to 0.80 P@50 (vs. 0.52 baseline) and 0.85 P@20 across held-out client domains.
• Engineered a reason-coded decision engine, deploying a reproducible end-to-end ML research paper and action playbook.

Research Intern, Advisor: Dr. Ayesha Hakim at KJ Somaiya School of Engineering (2026-07 – Present)

• Proposed and finalized the end-to-end architecture for a hybrid multimodal fake news detection framework integrating LLMs and Graph Neural Networks for evidence-first reasoning.
• Developing the system's core components, including dual-modality (text-image) encoders, a FAISS-backed RAG verification module, and a Heterogeneous Graph Attention Network (PyTorch Geometric) for cross-modal and propagation analysis.

AI/ML Intern at Elevate Labs (2025-06 – 2025-07)

• Engineered a deep Convolutional Neural Network (CNN) in Keras using Mel-Spectrogram audio representations, achieving 89% test accuracy across 10 distinct musical genres on 1,000+ tracks.
• Developed an automated spectral signal processing pipeline via Librosa to extract 20+ Mel-Frequency Cepstral Coefficients (MFCCs), spectral centroids, and zero-crossing rates, accelerating model training convergence by 25%.
• Analyzed class confusion matrices via custom diagnostics, boosting overall model F1-score by 12%.
• Built a low-latency inference engine for the.keras artifact, enabling sub-200ms audio predictions.

Bachelor of Technology in Electronics & Computer Engineering – KJ Somaiya School of Engineering (2024 – 2028)