Agent OS for financially regulated enterprises — inference pipelines, agent workflows, and production systems; initial domain debt collection.

Chapter organizer for the Dubai chapter of this global AI community.

Principal AI engineering consulting — architecture and delivery of applied ML and LLM-based systems.

Deepfake detection, manipulated-media analysis, phishing and social-engineering defence. BERT-based hate-speech classifiers; publisher verification workflows.

Co-founded a dating application built around an AI companion for finding meaningful relationships.

Co-founded a wellness venture.

DeepLearning.AI specialisations in GANs, LLMs, TensorFlow, and deepfakes; Google coursework in attention mechanisms, encoder-decoder architectures, and production ML systems.

Building Kyne — an agent OS for financially regulated enterprises. Designs agent workflows, inference pipelines, and production systems; debt collection is the initial regulated domain.

Dubai, UAE

Organizes the Dubai chapter of AI Salon — a global community of AI builders, researchers, investors, and partners.

Dubai, UAE

Principal AI engineering consulting — LLM and applied ML system architecture for client delivery in the post-2023 generative-AI cycle.

Dubai, UAE / Hybrid

Built detection systems for deepfakes, manipulated media, phishing, and social engineering. Developed PyTorch/BERT hate-speech classifiers and verification pipelines for publishers and institutions.

UAE

Led a team of six on APAC legacy-system migration and transformation. Built end-to-end ETL pipelines in PySpark on GCP (Dataproc, GCS layered storage, BigQuery views).

India

Automated credit risk scorecard development with Generalized Additive Models (GAM) and Partial Constrained Least Squares (PCLS) — 75% cost reduction (~$600K annually) versus vendor solutions. Trained LightGBM to filter 30% of credit-card applicants for straight-through processing with only 0.6% increase in bad rate.

India

Deep reinforcement learning for savings, budget, and investment nudges conditioned on risk appetite, financial goals, SMS data, and in-app behaviour. Resulted in three filed patents.

India

Led a team of four to build SMS extraction and classification using Support Vector Classification and Named Entity Recognition — accuracy improved from 72% to 91% over a rules-based baseline.

India

Designed a matching engine across 100+ UK banks and housing societies from KYC inputs and risk appetite, with daily web-scraped lending-criteria updates. Selected for UK FCA Regulatory Sandbox Cohort 4.

Bengaluru & London

Stress-test model selection via Monte Carlo tree search over feasible model sets — reduced average build time from 8 weeks to 0.5 weeks; winner of HSBC Inspiring New Innovation Award. Automated PRA Stress Test submissions in R/R Markdown (15 days → 2 hours).

Bengaluru

Traffic decongestion model with RNN/CNN and real-time feeds for Bengaluru City administration — 17% congestion reduction in a test area despite sparse data and irregular driving patterns.

Bengaluru

Multivariate time-series forecasting for day-ahead power demand and IEX auction prices — 40% accuracy improvement by incorporating weather variables. Mixed-integer linear programming unit-commitment model reduced solve time from 23 minutes to 3 minutes.

Kolkata

Part-number identification via document similarity (Jaccard, cosine, Euclidean distances). Promoter/detractor segmentation across 50,000+ survey responses and 2 GB of customer-care text using text mining on Net Promoter Score data.

Bengaluru

Brand and channel strategy for salons across North and Central India — 32% sector growth in six months.

Gurgaon

Graduate study in computer science.

Executive programme in business administration and management.

Coursework: Probability & Statistics, Time Series, Advanced Multivariate Analysis, Data Mining, Decision Models, Microeconomics, Optimisation, Operations Research, Operations Management. Master's thesis: Bayesian and frequentist event-study methodology for M&A risk–return analysis (implemented in R).

National graduate entrance examination in engineering and science.

Coursework: Linear Algebra, Calculus I–III, Algorithms, Databases, Total Quality Control, Reliability Management.

Formalises user engagement as a Markov-like state space with transition probabilities — from deep reinforcement learning work on financial nudges at Wizely.

IN 201941036721 · Ref: WZLY-301-INPR

Quantitative framework for measuring user response within a recommendation engine to evaluate personalised content policies.

IN 201941036731 · Ref: WZLY-302-INPR

Optimisation of content tone as a control variable for engagement — complements the reinforcement-learning nudge system at Wizely.

IN 201941036753 · Ref: WZLY-303-INPR

LLM methods and generative AI systems.

Production ML pipelines and deployment.

GAN architecture and training.

Neural networks, CNNs, sequence models.

Production TensorFlow development.

Deepfake detection and media forensics.

ML system design at scale.

Transformer foundations and sequence-to-sequence models.

Financial risk management.