Projects

Dreamers Inc sought to build a fully automated trading system capable of operating across multiple exchanges. The challenge: creating RL agents that could learn profitable strategies in highly stochastic, non-stationary financial markets — and then execute them live.

Led R&D from the ground up. Designed and built a custom in-house trading environment using Python and Cython for both simulation training and live execution. Explored and benchmarked advanced RL algorithms tailored to the partial observability and non-stationarity of financial data. Iterated rapidly with industry expert feedback to align models with real market dynamics.

Delivered a production-ready pipeline from training to live execution. The in-house environment became a versatile platform supporting rapid strategy iteration. Mentored a junior researcher who contributed to agent improvements.

Traditional credit scoring relies on bureau data, excluding large populations without formal credit history. Shield needed a way to assess creditworthiness for underbanked users using alternative signals.

Developed a credit scoring model leveraging mobile device data — app usage patterns, device metadata, and behavioral signals. Built a user profiling system that categorized individuals into archetypes (Businessman, Parent, Student, etc.) using sparse data recommendation techniques, enabling personalized risk assessment.

Expanded creditworthiness assessment beyond traditional metrics, enabling lending decisions for previously unscored populations. The profiling system became a cornerstone for personalized client services.

Fraudulent activities were causing significant losses. The company needed a robust, AI-driven system to detect and flag malicious behavior patterns from device-level signals.

Designed a fraud detection framework that analyzed device information patterns to identify anomalous and fraudulent behaviors. Combined feature engineering from alternative mobile data sources with classification models to flag high-risk users in real time.

Significantly reduced malicious activities across the platform. Enriched the company's analytics capabilities with newly identified alternative data sources from mobile devices.

Customer support agents were manually triaging hundreds of emails daily, leading to slow response times and inconsistent categorization.

Built an end-to-end email segmentation and classification pipeline using NLP techniques and SVM models. The system automatically categorized incoming emails by intent and urgency, routing them to the appropriate support queues.

Reduced email response time from support agents by 40%. The classification system handled the full volume of customer emails with minimal manual intervention.

The existing auto-debit system was monolithic, inflexible, and difficult to test or iterate on. Strategy changes required manual effort and were error-prone.

Single-handedly redesigned the entire auto-debit registration and initiation process, modularizing the system into composable, testable components. Built in the flexibility to rapidly test and deploy new collection strategies using A/B testing methodology.

Provided complete flexibility for strategy testing and implementation. Customer behaviour modelling on top of this system helped predict optimal calling dates, improving debt recovery rates by 30% for targeted user segments.

Surveillance systems need to track multiple objects simultaneously, even when objects temporarily occlude each other — a classic and challenging problem in computer vision.

Developed a multiple object tracking algorithm with specific handling for occlusion conditions. Implemented Kalman filters to predict the most probable trajectory of each moving object from detection signals. Deployed the full system for real-time inference on a Raspberry Pi.

Achieved real-time tracking of trajectory, velocity, and acceleration of multiple objects on edge hardware (Raspberry Pi), suitable for embedded surveillance applications.

Automating pick-and-place tasks on production lines requires robots to determine the optimal angle to grasp arbitrarily shaped objects — a non-trivial perception problem.

Trained a Convolutional Neural Network from scratch to predict optimal grasping angles for target objects. Used a sliding window technique for object localization on the TensorFlow framework. Integrated the model with a Dobot Magician robotic arm for end-to-end pick-and-place execution.

Successfully automated pick-and-place tasks with the robotic arm predicting and executing grasps based on CNN inference, demonstrating viability for production-line automation.