Ritesh Bansal Email and Phone Number

DART Solutions Engineering is a 40-person product engineering team of software engineers, quant developers, designers and product managers. The team's focus is on building quant and analytical platforms for risk modelers, risk and portfolio managers in Citi.DSE was established in 2021 and has designed, engineered and delivered the following platforms for Citi Risk:- Scenario Forecasting Platform (SFP) is at the core of Citi’s scenario creation process and is used by the Economic Scenario Team and its stakeholders to design, expand and review Scenarios.- Wholesale Forecasting Platform (WFP) is the next-generation platform for Citi's Wholesale stress testing and reserve calculations. WFP enables Citi’s model developers, analysts, sponsors, operations, and tech teams to collaborate on building, testing, and deploying CCAR/CECL/IFRS9 models.- Ratings Model Platform (RMP): RMP implements all DRM ratings models for Citi and makes them available via libraries and related analytical tools to model developers, risk managers and portfolio managers.In addition, I drive the following efforts around innovation and training:- Product Lab: A 12 week product management course centered around a capstone to taking a product idea from design to launch. - Innovation Lab: Incorporating new technologies, including GenAI, into products to improve productivity.- DART University: Provide upskilling pathways for various roles in Risk to increase the technical skill across the organization.

Risk Gears brought product thinking to risk infrastructure by pioneering Risk-as-a-Service (RaaS).Model-driven risk management became a regulatory-mandated and critical function in banking after the 2007 financial crisis. The pace of innovation in risk management technology platforms did not keep pace, leading to risk management costs spiraling from 3-5% in 2005 to 15% by 2017.Risk Gear’s RaaS product brought modern software infrastructure, component architectures, and cloud technologies to automate and streamline model development and deployment.

Safe Water brought market-based solutions for clean water to developing markets. My pro-bono work with Safe Water focused on establishing their digital fundraising strategy and bringing data-driven analytics to their water operations around the globe.

VerusAI applied industrial scale ML to large real estate data sets to create innovative products for real estate brokerages, appraisers, and lenders. VerusAI's suite of products included:- Valuation Suite - A set of automated valuation models and confidence scores for every residential property and historical values- Seller scores - Ranking of home sellers for lead generation to increase marketing spend efficiency by 3-5x. The data was used for direct marketing and targeting via customized audiences on Google and FB.- Loan Verity - Loan to Value updates for Fannie/Freddie issued agency pools.VerusAI created two technical innovations to rapidly innovate on complex real estate data:Data Fabric - Real Estate data comes in various formats: GeoSpatial, Timer Series, Panel and occasionally unstructured text. VerusAI’s Data Fabric integrated GeoSpatial data infrastructure across the 160 million parcels in the United States with 255 million historical sales and 237 million mortgage lien transactions over 30 years. The Data Fabric also included categorical and ordinal (ranked) datasets such as school rankings, tax rates, census data, income models, and finally flood and hazard models.Model Fabric: VerusAI achieved high model accuracy by creating localized models (e.g. county-level). This means VerusAI had to develop and maintain thousands of models across the 3144 counties in the US. The Model Fabric solved this problem by creating a Model Description Language (MDL) running on a compute layer. MDL allowed a model developer to succinctly describe a model along with derived features and the modeling layer automatically picks up the data elements for the county and runs it on the compute infrastructure.

Led a small team to reimplement the model infrastructure for stress testing one of the largest portfolios in the bank. Focused on bringing correctness, and verification to the model infrastructure while lowering developer time by moving towards a low-code solution.Highlights:- Model Expressions Engine (MEX): MEX automated model evaluation which led to reducing implementation time by over 80% and enabled rapid experimentation. MEX also led to elimination of over 90% of model implementation errors.- Model Unit Testing Templates (MUTT): Risk models are simulations and thereby hard to test and verify. MUTT embedded unit and system testing in the model implementation pipeline to greatly increase the quality and confidence in the resulting implementation.- Data Validation and Preprocessing (DVP): DVP performed data validation via rules-based set membership, distribution distance and range checks. Missing values were imputed in a rules based manner.- Performance primitives and parallelization - Library of performance primitives native to Numpy and Blas that sped up the model by over 10x. Parallelization used Dask library to distribute model blocks.

Data science and Machine LearningConsulting work in finance, marketing and politics.I worked at the intersection of data, machine learning, and economics. Classic statistical models are parsimonious and lead well to explanations. Machine Learning (ML) models use data (independent variables) profusely but are impossible to interpret. Statistical algorithms can have a couple hundred parameters but machine learning algorithms can have millions of features and hundreds of millions of rows. This allows for the use of very rich and sparse datasets at the expense of simplicity but results in higher accuracy. The financial domain poses unique challenges for the application of machine learning because of the time series nature of data. I have worked on adapting ML modes to financial time series prediction and worked extensively in data engineering for structured and unstructured data. This entails collecting, parsing, normalizing, entity resolution and loading. The work was done mostly in Python, Pandas, ML toolkits such as Scikit-learn, Vowpal Wabbit and datastores such as MySql, Postgres, Redis etc.

Quantitative strategies, High frequency trading, Index ArbitrageBuilt a high frequency trading group to deploy Index arb, ETF arb, along with intraday strategies. The technical objectives were to achieve a 10x cost reduction so we can could deploy strategies with lower payouts. The platform was built on a scrubbed down version of Linux with open source components for distributed filesystems, tick warehouses, computational libraries, sql and nosql database, and routing middlewares. I designed the platform, hired the technical team to develop the platform, and onboarded traders onto the platform. The platform included alpha testing and simulation infrastructure, trading infrastructure, and real-time risk management. The platform presented the quant trader a unified API to tick data warehouses, order books, order management apis, position management and strategy deployment (i.e. infrastructure virtualization).

Expanded MLP's U.S. arbitrage desk into emerging markets. The focus was on deploying Index Arb, corporate structure, closed end fund, and other arb strats in the markets of India, Taiwan, and Korea. Ported the U.S. quantitative models and implemented the automated execution system along with structuring MLP's offshore entities, acquiring regulatory licenses, negotiated clearing, and funding arrangements for emerging markets.