{ "personal_details": { "name": "Soumedhik Bharati", "email": "soumedhikbharati@gmail.com", "phone": "+91-8240947878", "linkedin": "https://www.linkedin.com/in/soumedhik-bharati-50b2bb203/", "github": "https://github.com/soumedhik", "location": "West Bengal, India" }, "education": [ { "university": "Sister Nivedita University", "degree": "B.Tech in Computer Science Engineering", "gpa": "8.68/10", "start_date": "Sept. 2022", "end_date": "Sept. 2026", "selected_courses": [ "Linear Algebra", "Probability", "Problem Solving", "Data Structures & Algorithms", "Object-Oriented Programming", "Operating Systems", "Database Management Systems", "Computer Architecture", "Discrete Mathematics", "Biology", "Automata", "Engineering Physics", "Digital Electronics" ] } ], "research_experience": [ { "title": "Research Assistant", "lab": "Xu Lab, Carnegie Mellon University", "start_date": "Sept 2025", "end_date": "Present", "responsibilities": [ "Engineered a few-shot transfer learning pipeline by fine-tuning a 1.2B parameter spatio-temporal Transformer, pre-trained on a 100,000+ hour clinical EEG corpus. Achieved a 12% absolute improvement in zero-shot seizure prediction, requiring 95% less labeled data than training from scratch.", "Implemented a self-supervised contrastive learning objective during pre-training, producing a 3x more robust latent space and reducing downstream fine-tuning convergence time by 80% on 5 distinct neurological datasets." ] }, { "title": "Research Assistant", "lab": "Sister Nivedita University", "start_date": "Mar 2024", "end_date": "Present", "responsibilities": [ "Designed HCAT-Net, a novel architecture for ordinal EEG emotion classification from 1D time-series data. The model integrates 1D ResNet blocks with a hierarchical Transformer encoder using Rotary Positional Embeddings (RoPE) and a cross-attention fusion layer to model multi-scale temporal dependencies. Introduced a custom Balanced Ordinal Loss function combining cross-entropy with a scaled MAE, achieving state-of-the-art results on the \"EEG Brainwave Dataset\" with a 99.8% test accuracy, 100% ROC-AUC, and a 99.7% MCC (presented at CIACON 2025).", "Proposed a novel DNA sequence encoding technique and integrated it into a hybrid CNN-BiLSTM architecture, achieving 97.2% classification accuracy across 6 classes while reducing memory overhead and model parameters by 35% compared to traditional one-hot baselines (under review).", "Implemented CADET, a BiLSTM-based essay evaluation architecture with multi-head attention, achieving state-of-the-art performance with QWK of 0.98, MSE of 2.88, and R² of 0.96 on the ASAP dataset (under review).", "Engineered a Reinforcement Learning agent to optimize employee training curricula de novo, without reliance on historical data. Formulated the problem as a POMDP and implemented an Actor-Critic agent with a hybrid reward function to solve the multi-objective task of maximizing skill gain under a hard budget constraint. The optimized policy achieved an 82% success rate and a budget overrun frequency of only 18% in a deterministic environment, outperforming four alternative reward-shaping strategies (manuscript in preparation).", "Developed and benchmarked a novel multi-scale UNet architecture for single-image dehazing, integrating Mamba state-space models and Ghost Convolutions which achieved a 55% reduction in trainable parameters and a significant decrease in computational load (GFLOPs) compared to baseline UNet architectures. It attained state-of-the-art (SOTA) performance on the RESIDE-6K benchmark dataset, while demonstrating competitive, near-SOTA results on challenging real-world haze datasets including O-HAZE, I-HAZE, and NH-HAZE, validating its efficiency and generalization capabilities (manuscript in preparation)." ] }, { "title": "Research Intern", "lab": "Indian Institute of Technology Kharagpur", "start_date": "May 2025", "end_date": "July 2025", "certification": "https://drive.google.com/file/d/1XtskkWOmgByOZCOxKppRbR9d5ERg99FY/view?usp=sharing", "responsibilities": [ "Developed a novel Information Retrieval (IR) reranking pipeline utilizing Large Language Models with a multi-stage caching mechanism to accelerate large-scale, reproducible experiments (manuscript in preparation).", "Engineered a novel parallelism strategy that outperformed top-down partitioning and sliding window approaches by 33% and 66% respectively in inference time, while maintaining or improving core IR evaluation metrics such as NDCG@k and MAP.", "Employed High-Performance Computing (HPC) infrastructure (Param Vidya cluster) using SLURM-based shell scripts for distributed training, inference, and large-scale benchmarking." ] }, { "title": "Collaborative Research", "lab": "University of Lille", "start_date": "Feb 2025", "end_date": "Present", "responsibilities": [ "Developed a deep learning surrogate model to rapidly predict the coherent evolution of a quantum system, bypassing direct numerical integration of the Schrödinger equation. The model maps a set of Hamiltonian control parameters and an initial wavefunction to the final evolved quantum state using a feedforward architecture with batch normalization and optimized dropout regularization.", "Achieved an R² of 0.94 against the ground-truth solver and demonstrated the superiority of a Huber loss function for handling heavy-tailed error distributions in the state space.", "Conducted in collaboration with postdoctoral researchers at the University of Lille (France); manuscript currently under review." ] } ], "work_experience": [ { "company": "Exalt.ai", "position": "Product Engineer", "start_date": "Jun. 2025", "end_date": "Present", "responsibilities": [ "Deployed a production-scale RAG pipeline for a high-traffic news summarization service, implementing a hybrid retrieval strategy fusing BM25 sparse lexical search with dense vectors from a Faiss index, and employing a ColBERT-style re-ranker to enhance contextual relevance and mitigate hallucination.", "Fine-tuned multiple large language models (LLMs) for domain-specific tasks, including news summarization and sentiment analysis, achieving a 22% improvement in summarization accuracy and a 15% reduction in model inference time through parameter-efficient fine-tuning (PEFT) techniques.", "Implemented model quantization and distillation techniques to optimize LLM performance for edge devices, resulting in a 40% reduction in model size while maintaining 90% of the original model's performance." ] }, { "company": "Raapid.ai", "position": "R&D Intern", "start_date": "Apr. 2025", "end_date": "Jun. 2025", "link": "https://drive.google.com/file/d/1oxNVFFB66f6LnJlW-KimO93e1-mfhyeV/view?usp=sharing", "responsibilities": [ "Developed a novel deep learning model for Hierarchical Condition Category (HCC) code extraction from unstructured clinical notes, improving accuracy by 12% over the existing baseline.", "Contributed to the enhancement of a proprietary knowledge graph by implementing an automated entity-linking module, increasing data consistency and coverage by 18%.", "Optimized data processing pipelines for large-scale medical records, reducing data ingestion and preprocessing times by 25% through parallel processing and optimized query design." ] } ], "projects": [ { "name": "Automatic Essay Grading System (SIT ICOE Hackathon Winner)", "link": "https://github.com/Soumedhik/Essay-Grading-System", "description": [ "Engineered a novel BiLSTM architecture with multi-head attention for automated essay evaluation.", "Developed custom hierarchical attention layers to improve model focus on essay coherence.", "Designed a domain-specific pre-processing pipeline for better data handling.", "Achieved state-of-the-art performance with MAE of 1.166 and QWK score of 0.674 on the ASAP dataset (n=12,978 essays).", "Outperformed previous SOTA models by 8.3% in coherence assessment metrics." ] }, { "name": "Image Enhancement using Autoencoders", "link": "https://github.com/Soumedhik/Image_Enhancement_Autoencoder", "description": [ "Architected a multi-scale convolutional autoencoder with sub-pixel convolution layers for single-image super-resolution.", "Incorporated skip connections for enhanced feature preservation.", "Engineered a novel hybrid perceptual loss function combining SSIM-based structural similarity metrics with deep feature representations.", "Achieved a 42.8% improvement in PSNR over traditional interpolation methods.", "Implemented advanced training optimizations, resulting in state-of-the-art performance (31.2 dB PSNR, 0.897 SSIM) on the DIV2K benchmark dataset." ] }, { "name": "Multi-Modal Face Tracking and Analysis System", "link": "https://github.com/Soumedhik/-Face_Tracking_VGG16", "description": [ "Created an end-to-end deep learning pipeline integrating VGG16 transfer learning with custom heads.", "Utilized Albumentations for data augmentation (15+ techniques) to enhance model generalization.", "Achieved real-time performance (30+ FPS) with localization accuracy of 95.2% (IoU > 0.5).", "Reduced model size by 47% through architecture optimization, maintaining efficiency." ] }, { "name": "Assistive System for Blind People (Intel OneAPI Hackathon Winner)", "link": "https://github.com/Soumedhik/Blind-Aid-Intel_OneApi_Hackathon", "description": [ "Built a multi-task computer vision system for visually impaired assistance using Intel OneAPI optimizations.", "Integrated YOLOv9 for real-time obstacle detection (98.3% accuracy) and MIDAS for depth estimation (MAE < 10cm).", "Used ResNet50 for Indian currency denomination recognition, achieving 99.4% accuracy.", "Achieved 25 FPS on edge devices with a 14.2MB model size and < 40ms latency, demonstrating a 92% success rate in real-world testing with 50 visually impaired users." ] }, { "name": "Image-to-Music Synthesis System", "link": "https://aurora-steel-theta.vercel.app/", "description": [ "Engineered a modular image-to-music synthesis pipeline by developing: (1) a multi-modal feature extractor that fuses 512-D semantic embeddings (Vision Transformer, CLIP) with quantitative color (K-means histograms), edge (Canny/Sobel), and texture (GLCM) statistics; (2) a cross-attention network to map visual features to latent musical parameters (key, mode, tempo); (3) diffusion-based generators for chord, rhythm, and melody synthesis, constrained by rule-based music theory for harmonic coherence; and (4) a neural vocoder-style synthesizer using residual upsampling and Butterworth filtering for high-fidelity waveform generation.", "Attained sub-second encoder latency on a GPU and an end-to-end generation time of under one second per second of audio; quantified output quality by achieving high tonal consistency (±2 semitones RMSE), precise rhythmic accuracy (±5 ms timing jitter), and strong emotional alignment (>0.85 Pearson correlation between image sentiment vectors and musical valence/arousal embeddings)." ] } ], "positions_of_responsibility": [ { "title": "Core Technical Team ML Lead", "organization": "Google Developer Group (GDG), SNU", "start_date": "Feb 2024", "end_date": "Present", "responsibilities": [ "Led advanced workshops on transformer architectures and attention mechanisms, training 100+ students in deep learning implementation.", "Developed comprehensive curriculum covering PyTorch, TensorFlow, and deep learning architectures." ] }, { "title": "Machine Learning Lead", "organization": "SKEPSIS", "link": "https://drive.google.com/file/d/1UZKZxhpG31odsKAuNjj5v3gTTt7aHlqa/view?usp=sharing", "start_date": "Oct 2023", "end_date": "Present", "responsibilities": [ "Led 5 research initiatives in NLP and computer vision, supervising teams of 4-6 undergraduate researchers.", "Designed and deployed a Book Recommendation System using collaborative filtering and k-means clustering, improving engagement by 20%.", "Mentored 60+ students across multiple machine learning projects, with 4 successful project completions." ] } ], "skills": { "specializations": [ "Natural Language Processing", "Computer Vision", "Time Series Forecasting" ], "programming_languages": [ "Python", "R", "C", "C++", "SQL" ], "frameworks_libraries": [ "TensorFlow", "Keras", "PyTorch", "scikit-learn", "Pandas", "NumPy", "Matplotlib", "Seaborn", "OpenCV", "SciPy", "Hugging Face" ], "research_tools": [ "LaTeX", "MATLAB", "Tableau", "Power BI", "Zotero", "Git/GitHub" ] }, "conferences": [ { "name": "9th International Conference on Data Management, Analytics & Innovation (ICDMAI)", "participation": "Participated and felicitated", "link": "https://www.icdmai.org/" }, { "name": "International Annual Meeting of the International Alliance of Skills Development for Belt and Road & BRICS Big Data and AI Working Committee", "location": "Chongqing, China" }, { "name": "IEEE International Conference on Computing, Intelligence and Application (CIACON 2025)", "certification": "https://drive.google.com/file/d/1GOTy4vHaCuUQwFbvp-PVhZOvczi5KLA2/view?usp=sharing", "location": "Durgapur, India", "presentation": "Presented paper HCAT-Net a Novel Hierarchical Cross-Attention Transformer Network with Enriched Balanced Ordinal Loss for EEG Emotion Classification (Record ID: 65473)" } ], "awards_achievements": [ { "award": "2nd prize in the BRICS International Vocational Skills Offline Competition 2024", "details": "Represented India in this prestigious competition, competing against 178 top international competitors.", "link": "https://drive.google.com/file/d/1HsaW5RM0NtZbDHeDhpHaFFTie7fcF9MT/view?usp=sharing" }, { "award": "Top 3 Teams Prize at the ICDMAI Offline Hackathon 2025", "details": "Competed against 1000+ teams in this prestigious event.", "link": "https://drive.google.com/file/d/18j2yZ4JHT0riiXQEnUUsTGt2IT2Ui5iX/view?usp=sharing" }, { "award": "Best Presenter Award at IEEE CIACON 2025", "details": "Recognized for outstanding presentation of research on HCAT-Net for EEG Emotion Classification.", "link": "https://drive.google.com/file/d/1UTnBZj2NDuY2NHVhq6uny9vZiJb220XW/view?usp=sharing" }, { "award": "1st place in the SAP ICOE Hackathon 2024", "details": "Competed against 400+ teams in this prestigious event.", "link": "https://drive.google.com/file/d/183WWQLWCE8gLdSGYh5p39Guk8YBhGh9i/view?usp=sharing" }, { "award": "Selected for the India Regional Bootcamp of Google Solution Challenge 2024", "link": "https://drive.google.com/file/d/1tsqE6t2-L7c3Z0igJgi8s2g_lO62HV-0/view?usp=sharing" }, { "award": "1st prize in the ML Mania Hackathon 2024", "details": "MCKV College of Engineering's Pragati 2k24.", "link": "https://drive.google.com/file/d/16kCkakwmPGdIWKRSvAjXs0ozcnRL3T7k/view?usp=sharing" }, { "award": "1st place in the Intel OneAPI Hackathon 2024", "link": "https://drive.google.com/file/d/1-cxl-n3j10O81z4tLQTjVvo0xV56Y1e5/view?usp=sharing" }, { "award": "Attended and was felicitated at the 9th International Conference on Data Management, Analytics & Innovation", "details": "Participated in discussions on advancements in data management and analytics.", "link": "https://www.icdmai.org/" }, { "award": "Attended the International Annual Meeting of the International Alliance of Skills Development for Belt and Road and the BRICS Big Data and AI Working Committee", "details": "in Chongqing, China – Participated in discussions on skills development and AI.", "link": "https://drive.google.com/file/d/1zdLkM9Eg9sV3kSBb0auouGKc7yV6DNrD/view?usp=sharing" } ] }