Fi Money

I have been working with Fi since July 2020, and am a part of the core team, building next-gen banking solution for India! Currently leading the effort on Data Science for the Ask.Fi team.

Building a complete search experience using Natural Language Understanding (NLU) to surface contextual search suggestions, and understand the user's submitted search query to produce results that are fast, frictionless & relevant.

NLU Engine: Structuring an incoming query to a parsed output by identifying query intent and the entities in the query, allowing the back-end service to populate sections of the search results screen accurately.

• Intent Classifier: Identify the query's intent using transfer learning to train a DistilBERT model from Huggingface's transformers library on a sequence classification task to help direct users to the appropriate functionality on the app.
• NER Tagger: Breaking down the user’s query into named entities using transfer learning to train a contextual BERT model using Huggingface’s for token classification.
• Similar Queries model: Finding the most semantically similar questions to a user's query from a predefined list using a pre-trained MPNet model to build embeddings.
• Productionisation involved the implementation of an end-to-end pipeline for training, experimentation & deployment using the MLflow ecosystem, an environment based setup, optimisation of inference execution time, adding logging (ELK), monitoring (prometheus) & instrumentation (rudder) for greater visibility into working & performance.

Suggestions: Show a suggested list of queries to the user typing into the search bar that capture their intention. Supervised two interns, split the problem statement into 3 models that were deployed using tflite on the client's device to allow for for speed:

• Word completion: Trained a character-based neural language model to complete a partially typed word%, for every character that the user types.
• Query Auto-complete: Trained a generative language model to predict \& suggest the completed query to the user% so as to minimise the effort on the user’s end.
• Similar Query Recommendations: Used BERT to find similar queries that rephrase the user's query more explicitly.

Building a model to categorise user's individual transactions that serves as an input into further downstream products and models such as search, insights and rewards

• Designed requirements and mechanism to identify payments to the same merchants across varying protocols & transaction methods.
• Predicting a label for the payer and payee of each transaction based on an internal category hierarchy using data points from real-time transaction info, historical payments, merchant information & properties, location-based information as well as user-provided labels & inputs.

• Responsible for driving projects, mentoring and helping fellow DS team members, and helping define standards for productionising DS models.
• Responsible for Data Science recruitment (including interview training) for Data Scientists, Business Analysts and ML Engineers for Fi.
• I regularly conduct knowledge-sharing sessions for the data team and the wider org on ongoing projects, ML models & algorithms, usage of MLflow as a model lifecycle tool, feature stores etc.