AI/ML Engineer with GenAI & Real-Time ML Expertise
Resume:
Bhopal, India
************@*****.***
Aditya Sneh
AI/ML Engineer
Portfolio
GitHub
Work Experience
Project Junior Research Fellow Aug 2024 - Jul 2025 Systems and Informatics Research Laboratory Bhopal, MP
• Built and deployed LogMe, a 24 7 Android (Kotlin) sensing platform that collected 250+ GB of real-world mobile sensor and usage data from 70+ participants over 14 days, backed by Apache/Python REST APIs and MySQL for reliable ingestion.
• Engineered notification-triggered labeling workflows capturing user activity, social context, intervention feasibility, and mood for supervised learning and adaptive real-time modeling.
• Developed and fine-tuned DySTAN (Dynamic Cross-Stitch, Cross-Task Attention, BiLSTM), achieving 0.852/0.876 macro-F1 with +21.8% improvement over a CNN-BiLSTM-GRU baseline.
• Built end-to-end ML pipelines (segmentation, downsampling) for production ML systems and implemented a contextual multi-armed bandit (Thompson Sampling) for real-time optimization, exceeding the 0.5 intervention reward baseline.
• Collaborated with an AIIMS psychiatrist to integrate domain expertise into adaptive intervention policies developed a Streamlit dashboard for real-time monitoring, model evaluation, and automated data validation. Gen AI Subject Matter Expert (SME) May 2024 - Jul 2024 Grades Buddy, Remote
• Delivered 40+ production-grade GenAI solutions, building and deploying LLM-based and generative models (GANs, diffusion) for support automation, document intelligence, healthcare support, and data-driven workflows.
• Built and fine-tuned transformer-based architectures (LLaMA 2/3, Mistral 7B, BERT) via Hugging Face, implementing sentiment analysis, contextual generation, and RAG pipelines achieving 85% task accuracy.
• Applied LoRA-based PEFT for efficient domain adaptation and improved response reliability in LLM systems, including medical chatbot deployments.
• Developed a real-time speech-based sentiment classification system using Whisper for transcription and PyTorch classifiers, supporting scalable customer audio analysis. Skills
Programming: Python, SQL
Machine Learning & Deep Learning: PyTorch, Transformer-based Architectures (Hugging Face), scikit-learn, NLP, Computer Vision, Multimodal Learning
LLM & Generative AI: Large Language Models (LLMs), Fine-Tuning, LoRA/PEFT, Retrieval-Augmented Generation
(RAG), OpenAI API, LangChain, LlamaIndex, Prompt Engineering, Agentic Systems Vector Search & Retrieval: FAISS, Pinecone, Chroma Backend & Data Systems: FastAPI, REST APIs, Pandas, NumPy Cloud & Infrastructure: AWS, GCP, Docker
MLOps & Production ML: Model Deployment, Inference Optimization, CI/CD, MLflow, Monitoring & Observability, MLOps Version Control: Git, GitHub
Education
BS-MS: Data Science & Engineering, Indian Institute of Science Education and Research (IISER), Bhopal 2019–2024 CPI: 8.06/10
Publications
• Aditya Sneh, A. Agarwal. On the Robustness of Iris Presentation Attack Detectors. BMVC 2025.
• Aditya Sneh, N. Sahu, A. Adyasha, A. Shelke, H. Lone. HCFSLN: A Hyperbolic Few-Shot Learning Framework for Anxiety Detection. arXiv 2025. Under review.
• Aditya Sneh, N. Sahu, H. Lone. DySTAN: Joint Modeling of Sedentary Activity and Social Context from Smartphone Sensors. arXiv 2025. Under review.
• A. Shelke, Aditya Sneh, A. Adyasha, H. Lone. Fairness-Aware Few-Shot Learning for Audio-Visual Stress Detection. arXiv 2025. Under review.
• N. Sahu, Aditya Sneh, H. Lone. Real-World Receptivity to Adaptive Mental Health Interventions: Findings from an In-the-Wild Study. arXiv 2025. Under review.
Projects
NotifSense: Real-Time Mobile Context Modeling & On-Device LLM Inference
• Built an on-device Android notification intelligence system in Kotlin that combines user context and notification metadata to make real-time ALLOW / DELAY / SUPPRESS decisions.
• Developed a PyTorch multi-label context model over 79 deployable mobile-sensor features reduced the label space from 24 to 9 and tuned per-label thresholds, achieving Macro F1 = 0.558 on held-out data.
• Optimized the HAR model for mobile with dynamic INT8 quantization, reducing size from 0.26 MB to 0.08 MB (3.1x smaller) while preserving on-device inference compatibility.
• Evaluated TinyLlama-1.1B for on-device reasoning using llama.cpp GGUF Q3/Q4 variants, enabling a privacy edge-LLM path under mobile memory constraints.
SensorFusionAgent: Agentic AI for Research-Grade Sensor Harmonization
• Built a local-first async Agentic-AI platform for multimodal sensor harmonization that infers schema/task context, aligns IMU streams, and produces fused outputs with transparent quality, drift, and confidence reports.
• Implemented a LangGraph Planner–Executor–Observer loop for safe pre-fusion optimization, applying unit scaling, axis inversion, and smoothing only when quality improved by 0.002.
• Designed HQScore v4, a 6-component fusion metric with explainable traces across KL, Wasserstein, FFT/spectral similarity, cross-correlation, normalized DTW, and drift stability.
• Added an adaptive ML layer with GradientBoostingRegressor models for sampling-rate and HQScore prediction, plus optional gpt-4o-mini fallback for schema inference on ambiguous inputs. Enterprise Decision Memory System (EDMS): Multimodal RAG Platform
• Engineered an enterprise RAG platform (FastAPI + React/Vite) for evidence-grounded Q&A over ADRs, RFCs, incident notes, postmortems, tickets, and architecture diagrams.
• Implemented a hybrid retrieval stack using text-embedding-3-small, FAISS, and BM25, with gpt-4o-mini generation constrained to retrieved enterprise evidence.
• Added multimodal ingestion with gpt-4o to convert uploaded architecture diagrams into retrievable text, plus JWT-based RBAC with SQLite-backed authentication.
• Developed an offline retrieval evaluation harness with Precision@K, Hit@K, and MRR over curated enterprise queries.
Contact this candidate