Bipul riding a fat-tire bike on a campus road

Senior Applied AI Scientist · Poshmark

Bipul Islam

I build production AI/ML systems. For 10+ years that meant crash detection and driver-scoring models for millions of drivers in mobility and telematics; now it's bringing my expertise in Behaviour, Trust and Safety to areas like marketplace operations, entity understanding and more. Before that, a PhD on control-based computer vision.
LinkedIn GitHub ORCID
10+ yrsbuilding production AI/ML
3.5 yrsin academia

Experience

  1. Poshmark

    Feb 2026 – present
    Senior Applied AI Scientist II
    Marketplace operationsEntity understanding

  2. Cambridge Mobile Telematics

    Mar 2021 – Jan 2026
    Senior Data Scientist
    • IoT-only usage-based insurance: led the launch of a new UBI offering with a novel server-side ML crash-processing framework on IoT sensor data, serving about 10 million end users.
    • Smart claims adjustment: led the design and release of a system that saves up to 40% of claims-adjuster time on in-network crash claims, for 7 million State Farm drivers.
    • Crash prediction: improved recall and reliability with a two-stage crash-detection model (a neural network plus LightGBM scoring), rolled out to 8 million users.
    • Edge computing: released an edge framework with a high-rate, two-stage phone model, including a CNN (PyTorch) for real-time crash detection, for 2 million users.
    • Total-loss prediction: released a framework saving customers $500+ per claim after a crash, for over 2 million users.
    • IoT hardware and firmware: led stakeholder buy-in for the hardware and firmware changes behind a Crash While Parked product targeting 15% of US insurance claim volume.

  3. Stony Brook University · SUNY Research Foundation

    Aug 2017 – Dec 2020
    Research Assistant, fully funded PhD scholar
    • Reformulated gradient-descent optimization in a feedback-control framework, improving computer-vision algorithms for 3D reconstruction and 2D–3D pose estimation.

  4. Zendrive

    Jul 2014 – Jul 2017
    Associate Data Scientist II
    • Real-time risk detection: deployed ML models that detect risky events during trips: hard braking, speeding, aggressive acceleration and turning.
    • Safety scoring: helped design and deploy a proprietary scoring method for continuously tracking driver safety.
    • System optimization: cut spurious trips on the server and saved up to 10% battery life on the client.
    • Data quality and monitoring: built a signal-processing pipeline in R for downstream teams, and a system-health platform on Elasticsearch with automated alerting.

Projects

Agentic Logistic Handler

Describe demand, supply and constraints in plain language; an agent identifies the right optimization problem, solves it exactly, and explains and maps the plan.

Google ADKGeminiPuLPFastAPI

LangChain Playground

A REST API and chat playground for a RAG assistant that role-plays any character, with an input guardrail and a second chain that tags each reply's mood.

LangChainLangServeFAISSOpenAI

Image Segmentation with Active Contours

Gradient-descent 2D active contours on level sets: Chan–Vese, Yezzi and Bhattacharyya functionals, plus interactive feedback control.

Computer visionPython

Automated Document Index Building

A framework comparing three NLP approaches to building a document's index: TF-IDF thresholds, LDA topic modeling, and Word2Vec with K-means.

NLPPython

More, from compilers to MapReduce, on GitHub.

Research

Doctoral thesis
Publications
Patent
Talks and mentoring

Skills

Languages
PythonRC/C++MATLAB
ML and data science
PyTorchscikit-learnLightGBMPandasNumPy
LLMs and generative AI
LangChainGoogle ADKGitHub Copilot
Big data
PySparkAzure Databricks
Cloud and databases
AWSPostgreSQL
Tools
GitDocker

Education

Beyond work