Selected projects.

Designed and led the development of a 360° Sales Intelligence capability that combines internal enterprise information with external web intelligence to provide sales teams with a more complete understanding of their accounts.

The platform builds comprehensive restaurant profiles covering digital presence, customer engagement, menu and product offerings, purchasing behavior, sales trends, and market positioning. AI-driven analysis identifies cross-sell, up-sell, retention, and account-growth opportunities and translates complex information into actionable recommendations for Sales Representatives.

The project deepened my understanding of how internal structured data, unstructured external information, LLMs, and autonomous agents can work together to create contextual intelligence that supports real-world sales decisions.

Led the development of computer vision solutions for banking workflows focused on cheque fraud detection and offline handwritten signature verification.

The work involved building and training deep neural networks capable of learning discriminative visual representations from cheque and signature images, allowing the systems to identify suspicious patterns and determine similarity between handwritten signatures.

Beyond model development, the project covered the broader machine learning lifecycle—from data collection and preparation through experimentation, training, evaluation, deployment, and production maintenance.

The project strengthened my understanding of applying deep learning to high-precision financial workflows where model performance, reliability, and production scalability are equally important.

Designed an Agentic AI–based solution for automatically interpreting balance and financial data and converting complex numerical information into concise, decision-ready business commentary.

The solution concept focused on moving beyond simple metric reporting by identifying significant movements, interpreting variances, connecting related financial signals, and generating contextual explanations that help business users understand what changed and where attention may be required.

The work strengthened my thinking around grounded AI-generated commentary, analytical reasoning over structured data, and the importance of combining deterministic financial calculations with generative capabilities rather than relying on an LLM for numerical reasoning alone.

Developed a deep learning–based OCR solution for extracting information from documents containing both machine-printed and handwritten text.

The project required treating document understanding as more than simple character recognition. Different document structures, handwriting variations, image quality, and visual layouts introduced challenges across preprocessing, text localization, recognition, and structured information extraction.

The work strengthened my understanding of document AI pipelines and how computer vision and text-processing techniques can be combined to transform unstructured document images into information that downstream enterprise systems can consume.

Led the development of an intelligent News Aggregator that combines traditional NLP and machine learning techniques with Generative AI to process and organize large volumes of news information.

The solution brought together different generations of NLP techniques, using conventional models where deterministic classification and extraction were effective while introducing generative capabilities for tasks requiring richer language understanding and interpretation.

The project reinforced an important architectural principle in my work: an effective AI system does not need to solve every problem with the same model. Combining specialized classical models with modern generative models can provide a better balance of accuracy, efficiency, explainability, and computational cost.

Built robust Name and Address Parsers for extracting structured information from noisy and inconsistently formatted textual data.

The solution addressed a common enterprise data challenge: semantically equivalent information can appear in many different textual forms because of formatting differences, abbreviations, missing fields, spelling variations, and other data-quality issues.

The project strengthened my understanding of information extraction, normalization, sequence-based NLP, and the importance of designing machine learning systems around imperfect real-world data rather than assuming clean and consistently formatted inputs.

Designed and led the enterprise deployment of an Agentic AI platform that ingests, validates, classifies, and analyzes procurement data to provide actionable intelligence across spend, supplier performance, sourcing opportunities, and procurement exceptions.

The solution uses intelligent orchestration and autonomous agents to identify cost-saving opportunities, supplier risks, and root causes behind procurement variances. Its architecture was designed for production-scale usage with caching and LLM optimization strategies that reduce latency, token consumption, and compute costs while maintaining enterprise security, governance, and scalability.

The project strengthened my experience in taking Agentic AI systems beyond prototypes and designing them around real enterprise workflows, operational constraints, and measurable business outcomes.

Built a similarity search engine for retrieving the most probable entities associated with transactions when the available information was noisy, incomplete, or inconsistently represented.

The system focused on matching imperfect transaction information against known entities and ranking potential matches according to similarity, allowing downstream workflows to work with probable entity associations instead of relying exclusively on exact matching.

The project significantly shaped my thinking about search and entity resolution. Real-world retrieval is rarely about finding an identical record; it is about representing information in a form where semantically or structurally similar entities can be discovered and ranked despite noise in the underlying data.

Led the development of the company's first automated smart surveillance system for a global investment organization, creating an event-driven computer vision platform capable of operating autonomously over visual information.

The system combined person tracking, face detection and recognition, office-space recognition, activity-area detection, and suspicious-activity detection to understand how individuals interacted with physical spaces. Visual inference was used to create unique customer profiles and derive activity insights that could improve customer support and situational awareness.

A central challenge was moving beyond isolated computer vision predictions. Detecting a face, person, or location provides only a momentary observation; meaningful intelligence emerges when those observations can be associated across time and transformed into events and persistent contextual information.

The project established much of my practical foundation in computer vision and helped shape how I approach AI systems today: not simply as individual models, but as coordinated pipelines where perception, state, events, and business logic work together to solve an end-to-end problem.

Designed an AI-enabled solution concept for creating a unified and governed source of sustainability performance information across organizational KPIs and supplier data.

The platform was designed to consolidate fragmented sustainability information, surface KPI and supplier performance, identify areas requiring attention, and use AI to generate recommendations for improving sustainability outcomes.

The project provided an opportunity to explore how Agentic AI can operate over governed enterprise data while preserving traceability and ensuring that generated recommendations remain grounded in measurable sustainability indicators.