Headline

Data Scientist

About

I build applied AI systems that translate human behavior into measurable signal — from psychometric modeling and speech-based emotion analysis to production-grade agentic pipelines and retrieval-augmented architectures. As a Data Scientist, I've spent the last couple of years designing and shipping systems at the intersection of behavioral science and applied ML: personality and organizational influence-inference engines, acoustic and linguistic scoring pipelines for emotional and communicative signal, and multi-agent LLM systems for automated data science workflows. My work spans the full stack — from statistical calibration and cross-validation methodology to API design and production deployment. Technical focus areas: - Generative AI & agentic systems (multi-agent orchestration, LLMOps) - Speech & audio ML (emotion recognition, synthetic voice generation, acoustic feature engineering) - NLP & psycholinguistics (lexicon-based inference, fine-tuned LLMs for classification tasks) - RAG architectures (multimodal, graph-based, and hybrid retrieval systems) - Applied statistics (calibration, cross-validation, large-scale benchmarking) Before this, I worked across healthcare AI, analytics/edtech, defense electronics (sonar-based signal processing), and R&D at a Fortune 500 hardware company — experience that's given me a strong foundation in both classical ML and modern deep learning approaches. I hold a B.Tech in Electronics & Communication Engineering and an M.Sc. in Data Science, and I'm continuing to build depth in GenAI and agentic system design. Open to conversations about applied AI, behavioral modeling, and production ML systems.

Location

India

AI stack

cursor

opencode

claudecode

Work experience

Data Scientist at ValueMatrix.ai, Sep 2025 - Present · 1 yr, Lewes, Delaware, United States

Data Scientist at WellWiz Innovation, Feb 2025 - Jun 2025 · 5 mos, Noida, Uttar Pradesh, India, At WellWiz Innovation, I worked across the full stack of applied health AI — from raw signal processing to agentic system design. I built a remote photoplethysmography (rPPG) pipeline for contactless physiological monitoring, extracting heart rate (1.5 bpm MAE), SpO2, and breathing rate directly from video — enabling multi-vital tracking without any wearable hardware. I developed multiple computer vision modules supporting this pipeline, including face detection and liveness/face-check verification to ensure signal quality and prevent spoofing, alongside preprocessing components for stable video-based signal extraction. I designed a multimodal health AI system using LangGraph, orchestrating multiple specialized agents to reason over heterogeneous health data (physiological signals, text, and contextual inputs) and produce coherent, clinically-relevant outputs rather than siloed single-model predictions. I also developed a mental health-focused RAG chatbot, grounding LLM responses in curated clinical/psychological knowledge bases to reduce hallucination risk in a sensitive domain. Separately, I built a CNN-based patient stratification model that classified patients into risk/outcome categories with 93% accuracy, supporting earlier and more targeted clinical decision-making. Across all five workstreams, the common thread was turning noisy, multimodal real-world health data — video, signals, text — into something a clinician or AI agent could actually act on.

Data Scientist at Analytics Vidhya, Oct 2022 - Sep 2023 · 1 yr, Gurugram, Haryana, India, At Analytics Vidhya, built an XGBoost-based malware classification system to detect malicious files from extracted feature sets, and developed a RAG-powered financial chatbot grounding LLM responses in financial documents/data for portfolio and market queries — combining classical ML for security with retrieval-augmented generation for finance.

Deputy Engineer at Bharat Electronics Limited, Feb 2021 - Jul 2022 · 1 yr 6 mos, Bengaluru, Karnataka, India, At Bharat Electronics Limited (BEL), I worked on sonar-based underwater surveillance systems supporting India's maritime defense capabilities, where accuracy and latency are operational necessities, not just metrics. I built models to detect and classify underwater objects from sonar signal data, distinguishing genuine targets from noise and multipath acoustic interference — reaching 95% detection accuracy through careful feature engineering on raw acoustic signals. I optimized the inference pipeline to run at 50ms latency, supporting real-time operational decision-making rather than after-the-fact analysis. Beyond the core detection models, I worked across the system's hardware-software stack: • Conducted EMI/EMC testing of sonar onboard electronics to ensure signal integrity and regulatory compliance in electrically noisy shipboard environments. • Administered Linux-based systems for the HMI (Human Machine Interface) and FDS (Fault Diagnostic System), supporting operator-facing controls and system health monitoring. • Developed ESI (External System Interface) modules in C++ to enable communication between the sonar system and external shipboard systems. • Built data simulators in C++ and Python to generate synthetic sonar/sensor data for testing and validation without dependency on live field data. • Developed and integrated firmware across multiple PCBs, handling low-level hardware-software interfacing for different subsystem boards. This combination meant operating under constraints uncommon in commercial ML work — strict reliability requirements, low tolerance for false negatives, and validation against adversarial acoustic conditions rather than clean benchmarks.

Data Scientist at HP, Jan 2020 - Jul 2020 · 7 mos, Bengaluru, Karnataka, India, At HP Inc., I built a federated learning system using LSTM networks and Autoencoders for anomaly detection in manufacturing processes — enabling model training across distributed data sources without centralizing sensitive production data. Beyond the modeling work, this involved designing the federated architecture itself — balancing local model updates with global aggregation — and tuning the LSTM to reliably distinguish genuine anomalies from normal process variance in time-series manufacturing data, without requiring raw data to leave individual sites.

Education

Bachelor of Technology, Electronics and Communication, Manipal institute of technology