Biswarup – Python, Machine learning, Pandas

The service’s main client, a connected-TV ad provider, relied on a legacy third-party identity solution to build lookalike audiences. Lovelytics’ managed audience recommendation service replaced it with two parts: a heterogeneous identity graph unifying 200M+ consumer profiles across multiple data providers, and a GraphSAGE-based model that expands an advertiser’s seed audience into a larger lookalike audience at household level. Live with the client since June 2025, it expanded their addressable reach from 21M to 85M households (~4x) while maintaining audience quality. It also saves them 500K+ a year, even after the cost of the managed service.

• Led ML R&D for the service end to end, from the models to the infrastructure.
• Architected the identity graph that resolves and unifies 200M+ consumer profiles across multiple data providers.
• Designed and trained the GraphSAGE-based model that generates the lookalike recommendations.
• Set up serving, with the models on A10 GPUs and the infrastructure on Databricks serverless compute.
• Result: replaced the client’s third-party solution, expanded reach ~4x, and saves the client 500K+ a year net of the service’s cost.

A Quebec hospital network wanted to run LLM-assisted workflows inside its own restricted data environment, relying less on Azure-hosted GPT services for reasons of cost, data residency and provincial health-data rules. As part of the engagement, the team turned clinician feedback into a specialized evaluation model. From 10,000+ clinician-assisted summaries, each clinician-edited summary was treated as preferred over its original draft. These pairs were used to adapt MedGemma 27B into a scalar reward model that scores clinical summaries against their source notes. The hospital now uses it to track the quality of summaries generated by smaller LLMs, reducing its dependency on frontier LLM APIs.

• Designed how preference pairs were derived from clinicians’ edits to generated summaries.
• Chose MedGemma 27B for its medical-text foundation and open weights, and designed the training and inference setup: LoRA on ~11K preference pairs, 2 epochs, about 8.5 hours on 2× A100 on Azure with NVIDIA’s training stack.
• Defined the success criteria and evaluation: held-out pairs split by patient and disease code to prevent leakage, plus a clinician-reviewed challenge set covering unsupported facts, omissions, negation, medication details and chronology.
• Worked on diagnosing the model learning shortcuts such as summary length and formatting, and addressing it by separating cosmetic edits from substantive ones, adding harder pairs matched for style and length, and selecting checkpoints on the challenge set.
• Result: 73% clinician-aligned pairwise accuracy on the challenge set, 12% more accurate than GPT-5. The hospital network now uses the model to monitor summary quality.

The product's goal was to help sales reps move more calls to the next stage, using each client’s historical call recordings to recommend what to do during live calls. As head of ML on the founding team, he built the ML infrastructure from the ground up and led the RAG-based agents behind the product. It scored thousands of recorded calls a day, ran multimodal RAG over call content, and used proprietary LLMs that were continuously improved through an internal labeling loop. In a 4-month pilot with Owner, one of the first customers, calls in the cohort using the recommendations converted to sales 6x more often than in the cohort that didn’t. The company was later acquired by Highspot.

• Created the ML infrastructure from the ground up to enable call scoring capabilities at scale; the infrastructure, hosted on AWS, scaled to thousands of calls per day.
• Spearheaded the development of RAG-based agents that powered the product.
• Built a retrieval system that was 2x more capable of finding quality call snippets than standard retrieval using OpenAI embeddings.
• Created internal tools for validation of results and automated reporting of unusual results.
• Created the continuous deployment pipeline.
• Brought seamless human-in-the-loop processes through customization of AWS SageMaker labeling jobs and gathered data for fine-tuning an in-house reranker model and LLM model.
• Fine-tuned an LLM for scoring calls based on annotated data using Mixtral 8-7b on Modal.
• Result: 6x higher sales conversion in the pilot cohort, and acquisition by Highspot.

A coffee brand client wanted personalized recommendations, including real-time recommendations at the drive-through. Biswarup built the recommender system, which generated over 1 million in incremental revenue compared with random assignment. He also built the experimentation platform used to evaluate models across customer cohorts.

• Built a recommender system for a coffee brand, generating over 1 million in incremental revenue over random assignment
• Developed an experimentation platform to evaluate ML models across customer cohorts, accelerating model iteration cycles.
• Optimized real-time inference for drive-through recommendations, reducing latency and deployment cost using ONNX RT in lambda
• Evangelized data-driven practices across teams via workshops and training sessions.

The hospital needed better data interoperability across its clinical systems and support for clinicians triaging COVID-19 patients. Biswarup implemented a next-generation clinical data repository on a FHIR server and built a COVID-19 severity prediction model to help clinicians with triage.

• Implemented a next-gen clinical data repository, using a FHIR server to enhance data interoperability.
• Built a COVID-19 severity prediction model to aid clinicians in triage.

A leading global provider of clinical research services, commercial insights, advanced healthcare analytics, and technology solutions for the life sciences and healthcare industries.

• Transitioned traditional analytics to a PySpark and Airflow-based automated analytics pipeline, cutting analysis time from 6+ hours to under 30 minutes.
• Developed an LSTM-based model to detect rare diseases from EHR data, which was integrated via a Flask-based API.