Projects

Industrial AI

Process Point Technologies · client work in mining and chemical processing

LIBS Spectral Calibration MLOps Pipeline · 2026 – Present

A laser spectroscopy (LIBS) analyzer predicts mineral concentration from spectral data, and its calibration model was being retrained by hand in notebooks. I designed an Azure ML pipeline that retrains the model on a fixed schedule (feature selection + ElasticNet), with a versioned database that tracks and audits every model across sample types. A scheduled scoring service writes predictions back into the plant’s process information system. A tiered hot/cold storage design keeps long-term data growth under control.

Stack: Azure ML SDK v2 · Azure Functions · Azure SQL · Blob/Parquet · OSIsoft PI

Mineral Beneficiation Process Optimization · 2026

Built models to predict valuable-mineral losses in a beneficiation (ore concentration) circuit, as inputs for a downstream process optimizer. I built data pipelines that combine process historian, laboratory, pump, flotation, and filtration data. Feature engineering covered stable-period detection, upstream concentration and throughput signals, and process lag measured through cross-correlation. I compared linear and tree-based models and documented the training workflows for handover.

Stack: Python · ElasticNet · XGBoost · Random Forest · Azure ML · OSIsoft PI

Predictive Pump-Failure Analytics Platform · 2025

Built an end-to-end predictive maintenance system for industrial slurry pumps that combines process sensor data with vibration telemetry. A configuration-driven deployment framework lets new assets be onboarded without code changes, with separate non-production and production endpoints. I also found and fixed a preprocessing bug in the legacy pipeline that had been discarding over 90% of training data.

Stack: Azure ML · OSIsoft PI Web API · Dynamox API · Python · ensemble regressors

Computer Vision for Haul Truck Monitoring · 2026

Proof-of-concept systems that use existing CCTV feeds to detect dump trucks and classify whether they are loaded or empty. I evaluated object detection and segmentation models against vision-language models for the classification step.

Stack: RT-DETRv2 · SAM · Qwen2.5-VL · SmolVLM · Python

Root-Cause Analysis for Fertilizer Plants

Applied multivariate, time-lagged analysis to sensor data from urea and ANP plants to trace the process conditions behind production losses, and turned the findings into recommendations operators could act on.

Stack: Python · time-series analysis · anomaly detection

VLM-Assisted Data Annotation for PPE Detection

Used vision-language models to generate annotations for a custom PPE detection dataset in-house, avoiding both manual labeling and public annotation tools for confidential images. Read the write-up →


Earlier Work

Drishya AI Labs · Data Scientist · 2022 – 2024

Computer vision and document understanding systems, including OCR pipelines with LLM-based post-processing to extract structured data from documents.

Deloitte Shared Services India · Data Scientist · 2019 – 2022

  • Machine Reading Comprehension chatbot: added transformer-based document question answering to Deloitte’s internal employee chatbot (2020).
  • Employee attrition prediction: predicted attrition from 40+ employee factors, using model interpretability methods to explain what drove the predictions.
  • Price quoting insights: gave project managers insights for pricing client engagements, based on project scope, staffing, and duration.

Academic & Open Source

  • Human Action Recognition (M.Tech thesis, 2019): real-time recognition of human actions from CCTV footage. GitHub →
  • Sukanya Rakshak (Junior Research Fellowship, 2018–19): a WSN/IoT security system for hostels, funded by CST Uttar Pradesh. GitHub →
  • Hinglish Language Identifier: RNN model in Keras that classifies words in Hindi–English code-mixed text. GitHub →
  • Abstractive Text Summarization: seq2seq model trained on 500K+ Amazon food reviews. GitHub →
  • Generative Chatbot (ChatCrazie): a context-aware sequence-to-sequence chatbot. GitHub →

Personal Projects

AI Email Agent

An n8n workflow agent that automates Outlook email handling.

Stack: n8n · LLMs · Microsoft Outlook