Work

Vertical LLM & GenAI solutions

Industry-specific AI applications that leverage large language models to solve domain challenges like legal document analysis or medical diagnosis assistance.

Core ML/DL Implementation

Fundamental machine learning and deep learning development involving algorithm selection, model architecture design, training pipelines, and optimization techniques.

Cloud Modernization and ML Infra

Upgrading existing cloud and infrastructure workflows to improve scale and reliability, Building infrastructure for ML operations including model deployment, scaling, and monitoring.

Software Product Engineering

End-to-end development of software products or features from concept to delivery, encompassing design, coding, testing, and continuous improvement.

Vertical LLM & GenAI solutions

Industry-specific AI applications that leverage large language models to solve domain challenges like legal document analysis or medical diagnosis assistance.

Dscout

UX Research

A qualitative UX research platform designed to facilitate remote, mobile-based, and in-the-moment research. It enables companies to gather insights into user experiences.

Scope: LLM assisted research analysis and development of new features using LLMs.
• Successfully delivered new features like Theme Extraction, Summary Generation, Live Video summarization, AI moderated interviews etc. using OpenAI and RAG based Q&A on specific studies.
• Successfully implemented LLM-based retrospective annotation across all features, significantly enhancing analysis workflow efficiency and accuracy.
• The enhanced capabilities delivered by the team were deployed to production, serving approximately 50 pilot customers.
‍
ML & Infra Frameworks: OpenAI APIs Instruction Finetuning Prompt Engineering LoRA DeepSpeed Huggingface Accelerate Llama-Index AWS Sagemaker Data dog Lamini Falcon Gemini

Advinow

Clinical Workflow Automation

Medical History Generation, AI-Powered Automated Narrative Creation for Clinical Documentation Using LLM.

Scope: Building a medical document QA system using RAG architecture and Gemini LLM.
• Developed an AI-powered HPI generation engine using GPT that automates comprehensive medical history documentation, reducing physician administrative time by 40% while maintaining high clinical accuracy and narrative quality.
• Improved patient experience by eliminating repetitive questioning, as the system intelligently synthesizes information already collected through the Advinow platform.
• Achieved a 95% acceptance rate from physicians for AI-generated HPI narratives, demonstrating high accuracy and clinical relevance of the generated content.
aGPT-4, Neo4J, Langchain, ElasticSearch, MongoDB, Prompt Engineering, FastAPI
ML & Infra Frameworks: GPT-4 Neo4J Langchain ElasticSearch MongoDB Prompt Engineering FastAPI

Giotto AI

Medical Document Processing

Systematic literature review for medical documents.

Scope: Building a medical document QA system using RAG architecture and Gemini LLM.
• Developed a Retrieval-Augmented Generation (RAG) system using Gemini LLM to create an advanced medical document question-answering platform, focusing on comprehensive document parsing including scanned PDFs, tables, and precise answer extraction through bounding box localization.
• Designed a robust cloud-based infrastructure using Google Cloud Platform (GCP), Docker for containerization, and Grafana for monitoring, ensuring scalable, high-performance deployment with comprehensive logging and error handling.
• Achieved a minimum 10% performance improvement over traditional BERT QA models, focusing on reducing manual document review time, enhancing medical research efficiency, and providing interpretable AI-driven insights for medical professionals.

ML & Infra Frameworks: Gemini APIs Vector Store Langchain Huggingface Bert Redis PubSub Message Queue Spacy NLTK Lamini PaddleOCR TesseractOCR

Zenius

Intelligent Tutoring Systems

An E-learning and online tutoring platform powered by cutting edge AI techniques to address various challenges in the education industry. Major part of the product includes an automated query resolution system that leverages vector search and deep learning technologies to provide efficient and accurate query responses.

Scope: Engineered a sophisticated vector search system utilizing FAISS (Facebook AI Similarity Search) that:
• Processes and vectorizes queries using BERT embeddings
• Performs efficient similarity search across large-scale educational content
• Delivers automated and accurate query responses

Designed and implemented a robust CI/CD pipeline for the vector search system:
• Automated deployment workflows
• Containerized solution using Docker
• Integrated with cloud infrastructure

Built a scalable API service using FastAPI to handle query processing and response generation. Also, utilized natural language processing with Spacy for enhanced query understanding and processing. FAISS, Docker, FastAPI, Google Cloud Platform, BERT, Spacy
Tech Stack: FAISS Docker FastAPI GCP BERT Spacy

Core ML/DL Implementation

Fundamental machine learning and deep learning development involving algorithm selection, model architecture design, training pipelines, and optimization techniques.

Recognic AI

Document AI

Cloud-native SaaS-based Document AI product that handles entire document processing lifecycle from ingestion to archive.

Scope: Building, Deploying and Managing ML models with various capabilities for the product that involved classifying, processing, parsing and extracting insights from documents at scale.

Cloud Services Used: GCP Stack (GCS, App Engine, Dataflow, Data Studio, Bigquery, VertexAI, AutoML, Google Vision OCR Engine)

ML Frameworks: YoloV5 EfficientNet CascadeTabNet Graph Convolutional Networks Siamese Networks FastText LayoutLM

Threose

Computational Genomics

Research project aimed at augmenting the aptamer discovery using AI and Deep Learning; by using and generating data from computational genomics labs.

Scope: Building extremely scalable ETL infra to process massive sequencing data from various biological experiments (viz. SELEX, PCR, BLI, NGS) using protocol buffers to maintain metadata and track the modelling and experimental lineages.
• Building Deep Learning algorithms to perform Generative and Predictive modelling for aptamers based on various signals derived from biological experiments.
• Generative modelling: Coming up with new sequences for a given biomolecular target.
• Predictive modelling: Classifying binders/non-binders
• Built and improved the Genetic Algorithm for generative modelling. Handled the GA infrastructure to parallely execute the GA search of sequences in a massive search space.

ML & Infra Frameworks: Diffusion Models BERT Apache Beam (Dataflow) Protocol Buffers GPT-2 Seq2Seq SAVAE Cloud Spanner Data Studio Data Proc

Embibe

Educational Content Analytics

An automated content analysis and metadata tagging system for academic materials, leveraging advanced NLP and machine learning techniques to enhance educational content organization and knowledge tracing.

Scope:
Engineered a comprehensive metadata tagging system that:
• Integrates with knowledge graph nodes for academic content classification
• Processes and tags YouTube video content automatically
• Implements image metadata tagging using deep learning techniques, achieving 12% improvement in F1 score
‍
Developed and fine-tuned transformer models for concept classification across a vast knowledge base:
• Covered 25,000+ academic concepts
• Achieved 59.34% accuracy in concept tagging
• Enhanced Knowledge Tracing algorithms resulting in 12% AUC score improvement
‍
ML & Infra Frameworks: Transformer Models Gradient Boosted Decision Trees GCP Docker Cloud Run Git Deep Learning Frameworks

Cloud Modernization and ML Infra

Upgrading existing cloud and infrastructure workflows to improve scale and reliability, Building infrastructure for ML operations including model deployment, scaling, and monitoring.

BlueUrbn

Estate Energy Optimization

An AI-powered platform that leverages machine learning and physics-based models to identify optimal energy optimization opportunities and retrofit recommendations for commercial and residential buildings, maximizing ROI and energy efficiency.

Scope: Architected and implemented a user-friendly dashboard that generates comprehensive retrofit recommendations based solely on building address input, integrating seamlessly with Google Maps JS API for location services.
• Designed and built a highly scalable distributed computing infrastructure to handle 1 million+ EnergyPlus simulations, utilizing:
      • Google Cloud Pub/Sub for message queuing and event handling
        • GCP Dataflow for parallel processing across hundreds of machine instances
        • Continuous 24/7 operation with robust job logging and monitoring
‍• Developed enterprise-scale ML model deployment infrastructure using Vertex AI, including:
      • RESTful API implementation for model inference
      • Integration with building analysis workflows
      • Scalable serving architecture for handling multiple concurrent requests
• Created end-to-end data pipelines to process and analyze building energy consumption patterns, enabling accurate and actionable retrofit recommendations

Tech Stacks: NextJS Python GCP BigQuery GCP Pub/Sub GCP Dataflow Git Google Maps API ReactJS NodeJS Express PostgreSQL Vertex AI Tailwind

Tapestry (GridOps)

Energy Grid Management

Gridops is an advanced operational support tool that optimizes electric grid management through economic dispatch solutions. It provides grid operators with actionable insights for day-ahead planning for distribution operations by forecasting generation, and consumption, thereby enabling smarter decision-making in meeting power demand.

Scope:
• Handled the complete modelling and infrastructure for load forecasting and generation forecasting from different modalities (SCADA Data and Meters Billing Data) for Chilean Energy providers using various Deep learning algorithms.
• Built end-to-end infrastructure to support periodic ingestion of weather covariates (from multiple weather API providers), and building the product (frontend dashboards and backend ETL pipelines) that showcases dispatch proportions across grid, and consumes forecasts to display and compare forecasting projections on the tool.
• Worked closely with CEN to achieve significant milestones on product POC by closely understanding the electric grid topology, mappings, data sources and its representations.
• Delivered the POC which secured a 3M$ contract for the client.

ML & Infra Frameworks: Timeseries Forecasting Unit8 Darts gRPC Apache Beam Dataflow Protocol Buffers Terraform AngularJS Bazel Protocol Buffers Vertex AI BigQuery Docker Cloud Run

Fractal

Supply Chain Optimization

A Python-based supply chain simulation system designed for MNCs to optimize their global supply chain operations by predicting and preventing stockouts while maintaining optimal inventory levels.

Scope: Developed a comprehensive data ingestion and processing pipeline to handle terabyte-scale client data from Google Cloud Platform, ensuring efficient data flow for simulation operations.
Engineered a sophisticated simulation environment that accurately models:
• Client-specific business rules and domain constraints
• Complex supply chain dynamics and operational workflows
• Load testing scenarios to identify potential bottlenecks
• Weak points in the supply chain that might fail under stress

• Built predictive modeling capabilities to forecast inventory requirements and optimize stock levels across all supply chain nodes, reducing the risk of stockouts while maintaining efficient inventory management.
• Implemented scenario analysis functionality to test various load conditions and assess supply chain resilience under different operational conditions

Tech Stack: gRPC Python Protocol Buffers GCP BigQuery GCS Bucket GCS Bucket Apache Beam

HappierWork

HR Management

Cloud-first HRM platform offering comprehensive workforce management capabilities in one integrated solution.

Scope: Building the resume parser and adding skill extraction capabilities to the hiring module (before the LLM era). Candidate-to-role matching and fitment scoring.

ML Frameworks: SpaCy NLTK Word2Vec FastText Siamese Networks TF-IDF vectorization Doc2Vec Elasticsearch XGBoost

Software Product Engineering

End-to-end development of software products or features from concept to delivery, encompassing design, coding, testing, and continuous improvement.

Omnisd

Industrial IoT

A monitoring and management system for ONGC's oil and gas wells in remote locations, enabling site operators to track performance and receive alerts for anomalies.

Scope: Develop comprehensive data ingestion infrastructure to process and store sensor data from remote oil and gas wells, ensuring reliable data collection and storage in PostgreSQL database.
• Built an automated anomaly detection system using scheduled SQL procedures to analyze daily sensor data, generate analytics, and trigger alerts for any detected abnormalities in well operations.
• Designed and implemented an interactive dashboard using ReactJS, allowing operators to:
       • Visualize real-time sensor data from individual sites
        • Monitor overall status across all operational sites
        • Access generated analytics and anomaly alerts
‍• Implemented containerized deployment architecture using Docker and Docker Compose, ensuring consistent and reliable deployment of the frontend, backend, and database components.

Tech Stacks: NodeJS NextJS ReactJS PostgreSQL Express Docker Tailwind Docker Compose Graphana Prometheus

Tectio

Smart HVAC

AI driven smart HVAC servicing and maintenance software tool.

Scope: Developed and managed dashboard which ingested the data from HVAC(IoT) sensors and generated complex visualizations to make the service technician's job easy by alerts generation and monitoring at customer site and equipment level insights.
• Visualization via complex data-heavy plots
• Customized visual architecture in D3.js to track air-flow & temperature the AC/Heater to easily figure out what part of system is non-functional.
• Created service technician's navigation flow on the dashboard which allows them to find site and equipment via Google Maps for easier filtering.
‍
Tech Stack: Vue.js React.js D3.js Docker React Query Websockets