Aishwarya Patange

Machinelearning,made inNew York

Data scientist at Gartner. Built an AutoML platform from scratch, trained detectors on live city intersections at Columbia, and shipped LLM systems for cybersecurity. I like the whole pipeline, from messy data to something people use.

See the workSay hello

Open to conversations · 2026

Aishwarya Patange

Five years of turning messy data into things people use

I started in Mumbai as an electronics engineer who spent more time training models than soldering boards.

  • 5+years building ML
  • 4cities, two continents
  • 12+open-source projects

That turned into building torsha.ai, an end-to-end AutoML platform, at iQGateway in Bangalore. Then a master’s at Columbia, where I trained detection and tracking models on real intersection footage for the COSMOS smart-city testbed and taught deep learning as a TA. Today I do data science at Gartner, working out of the Stamford office and still living in New York.

What I care about: models that hold up outside the notebook. Clean data pipelines, honest evaluation, explainability, and the engineering that gets a model in front of people.

Four cities, one direction

World map showing the route from Mumbai to Bangalore to New York, with Stamford as the current office

  1. Mumbai, India · 2017 – 2021

    Where it started

    B.E. in Electronics & Telecommunication at the University of Mumbai, top 10% of the class. Built a wearable assistant for the visually impaired and realized I would rather train models than tune circuits.

  2. Bangalore, India · 2021 – 2022

    First production ML

    Associate Data Scientist at iQGateway. Led the design of torsha.ai, an end-to-end AutoML platform, and learned what it takes to get a model in front of paying users.

  3. New York, USA · 2022 – 2023

    Columbia & COSMOS Labs

    M.S. in Electrical Engineering, ML & Data Science track. Trained YOLOv8 detectors on live intersection feeds, cut streaming latency to 80 ms, and TA’d Neural Networks & Deep Learning.

  4. Stamford, USA · 2024 – now

    Gartner

    Data science at Gartner, with the office in Stamford and home still in New York. Retention analytics: probabilistic account scoring across 10,000+ enterprise clients, and the evaluation framework behind an internal RAG assistant.

Where the work happened

  1. Jun 2024 – Present · Stamford, CT · based in New York

    Data Science at Gartner

    Retention and growth analytics for Gartner’s enterprise client base, owning the models that tell account teams which relationships need attention and why.

    • Architected and launched the Implied Value Rating model, an account-level probabilistic score covering 10,000+ enterprise clients at 94% accuracy, now in production triggering proactive retention action and contributing to a $295M revenue opportunity.
    • Engineered the feature pipeline behind the customer segmentation and propensity models, combining behavioral engagement signals with MiniLM topical embeddings into cohort-level features that lifted AUC by 7% over an unsegmented baseline.
    • Partnered with Sales Analytics on next-best-program targeting and recommendation models, batch-scoring 134.5K prospects and projecting $40M in savings from discontinuing low-performing programs.
    • Stitched two disparate internal sources, Surveys and Case Studies, into a single unified layer, owning the path from raw documents through the extraction pipeline to the OpenSearch indexes served at query time.
    • Built the evaluation framework for Research Desk, an internal RAG assistant, from scratch on Langfuse, pairing LLM-as-a-Judge checks for numeric grounding, fact attribution, and named-entity correctness with Precision@K and Recall@K per query category, lifting retrieval pass rate from 43% to 78% and factual accuracy by 13%.
    94%IVR prediction accuracy
    $295Mrevenue opportunity
    78%RAG retrieval pass rate, from 43%

    Python, MiniLM, OpenSearch, Langfuse, RAG, LLM-as-a-Judge

  2. Jan 2023 – Dec 2023 · New York, NY

    Computer Vision Research at COSMOS Labs, Columbia University

    Computer vision for the Smart City Intersections project on the NSF COSMOS testbed, under Prof. Zoran Kostic.

    • Trained and fine-tuned YOLOv8 models to detect and track vehicles, pedestrians, and two-wheelers from live intersection cameras, reaching 0.79 mAP.
    • Built a semi-supervised, model-assisted data engine inspired by Segment Anything to automate data collection for tracking, lifting model performance by 20%.
    • Designed custom TCP and UDP streaming protocols for the testbed, cutting video streaming latency to 80 ms.
    • Represented Columbia in the NSF Center for Smart Streetscapes, an Engineering Research Center.
    0.79mAP, detection & tracking
    +20%model performance
    80 msstreaming latency

    PyTorch, YOLOv8, OpenCV, Python, TCP/UDP

  3. Jun 2023 – Sep 2023 · New York, NY

    Software Development at ChaosTrack Inc.

    LLM-powered incident-response simulation for cybersecurity teams.

    • Shipped three major product features across a Node.js, REST, GraphQL, Hasura, ArangoDB, PostgreSQL, and Docker stack.
    • Built an LLM application that simulates and scores corporate incident responses, using spaCy, NLTK, PyTorch, and Hugging Face.
    • Used LangChain and LangSmith for prompt engineering, improving GPT responsiveness by 35% and output accuracy by 25%.
    • Ran retrieval-augmented generation experiments with Milvus and semantic search with OWL-ViT and Wav2CLIP embeddings.
    +35%model responsiveness
    +25%output accuracy
    3features shipped

    LangChain, Hugging Face, Milvus, Node.js, GraphQL, Docker

  4. May 2021 – Jul 2022 · Bangalore, India

    Machine Learning Engineering at iQGateway

    Core developer of torsha.ai, an end-to-end AutoML and data science platform.

    • Led the design and development of torsha.ai, an end-to-end data science platform with AutoML, model explainability, and custom DAG-based pipelines for training and evaluation.
    • Built the platform’s data visualization module with Plotly and D3.js.
    • Raised AutoML pipeline prediction accuracy by 15% with custom scikit-learn-compatible estimators.
    +15%AutoML accuracy

    Python, scikit-learn, AutoML, Plotly, D3.js

Things I built, and what they did

AutoML platform

2021 – 2022

torsha.ai

Problem
Data teams spend most of their time on plumbing: cleaning, feature pipelines, model selection, and explaining results to stakeholders.
Approach
Designed the core of an end-to-end platform: AutoML search, custom DAG-based pipelines for training and evaluation, scikit-learn-compatible estimators, and an explainability layer.
Outcome
15% higher prediction accuracy over the baseline pipeline, plus a Plotly and D3.js visualization module that became the product’s front door.
+15%prediction accuracy

Python, scikit-learn, AutoML, Explainability, Plotly, D3.js

LLM system · ChaosTrack

2023

Incident Response Simulator

Problem
Organizations rehearse cyber incidents rarely, and when they do, the drills are scripted and forgettable.
Approach
An asynchronous, chatbot-based simulator that plays out real-world incidents and evaluates each team’s response. Built on LangChain and LangSmith with Hugging Face models and retrieval over Milvus.
Outcome
Prompt engineering and retrieval lifted model responsiveness by 35% and output accuracy by 25%, setting a new readiness standard for customers.
+35%responsiveness
+25%accuracy

LangChain, LangSmith, Hugging Face, Milvus, RAG, Node.js

Open source, three picks

All repositories on GitHub ↗

Streaming ML

2023

Real-time stock prediction

Sentiment analysis on 5M+ daily tweets streamed through Pub/Sub, Apache Beam, and Dataflow into BigQuery, improving price prediction by 12% over the baseline.

+12%prediction lift
5M+tweets / day

Python, Apache Beam, BigQuery, Google Cloud, NLP

LLM · Retrieval

2024

LLaMA 2 with RAG

Retrieval-augmented generation with LLaMA 2 running fully on a local machine, backed by a vector database for grounded answers over private documents.

LLaMA 2, RAG, Vector DB, Python

NLP · From first principles

2024

Transformer from scratch

English to Italian neural machine translation with a 6-layer encoder–decoder transformer written from scratch in PyTorch. Trained on OPUS Books with TensorBoard monitoring.

PyTorch, Transformers, Attention, Hugging Face Datasets

The pipeline I ship with

No proficiency bars. These are the tools behind the work above, grouped the way a project actually flows. Anything with a red mark has shipped in something on this page.

Data & Infra

Getting data in, at scale, reliably.

  • Apache Spark
  • Apache Beam
  • Dataflow
  • BigQuery
  • Pub/Sub
  • PostgreSQL
  • MongoDB
  • ArangoDB
  • Docker
  • Google Cloud
  • Linux
  • Git

Modeling

Training, evaluating, and explaining models.

  • Python
  • PyTorch
  • TensorFlow
  • scikit-learn
  • Hugging Face
  • YOLOv8
  • OpenCV
  • AutoML
  • SHAP / Explainability
  • R

Retrieval & LLMs

Grounding language models in real data.

  • LangChain
  • LangSmith
  • RAG
  • Milvus
  • CLIP / OWL-ViT
  • OpenAI API
  • LLaMA 2
  • Prompt engineering
  • Fine-tuning

Serving & Apps

Putting models in front of people.

  • Node.js
  • Express
  • GraphQL
  • Hasura
  • REST APIs
  • JavaScript
  • Java
  • Streamlit
  • Plotly / D3.js
  • TCP / UDP streaming

Two degrees, one thread

Sep 2022 – Dec 2023 · New York, NY

Columbia University

M.S. in Electrical Engineering · Machine Learning & Data Science track

  • Graduate Teaching Assistant for E4040 Neural Networks & Deep Learning.
  • Research with Prof. Zoran Kostic on the Smart City Intersections project (COSMOS Labs).
  • Columbia representative in the NSF Center for Smart Streetscapes.
  • Coursework: Statistical Learning, Natural Language Processing, Sparse Models for High-Dimensional Data, Algorithms for Data Science, Big Data Analytics.

Aug 2017 – May 2021 · Mumbai, India

University of Mumbai

B.E. in Electronics & Telecommunication

  • Ranked in the top 10% of the class.
  • Built Blind Person’s Assistant, a wearable that uses face detection and recognition to help visually impaired users.
  • Coursework: Big Data Analytics, Operations Research, Image Processing & Machine Vision, Data Compression & Encryption.

What I’m up to lately

Building
An agent that documents my lifePhone, laptop, and speech capture stitched into a searchable diary. Retrieval again, on far messier data than work.
Reading
HyperionDan Simmons. Six pilgrims, one Shrike, and a lot of pages left.
Listening
The Strokes, on repeatWith a steady diet of Desi hip hop when the volume needs to go up.
Watching
Whatever is playingBig movie buff. Strong opinions, held loosely, offered freely.
Following
Manchester City & the KnicksOne of these teams causes more stress than the other.
Playing
Chess, and guitarOpenings I know, endgames I am learning. Riffs over scales.

Updated November 2026

Let’s build somethingthat ships

Open to conversations about data science and ML engineering roles, collaborations, or a good chess game. My inbox is open.

Say helloLinkedInGitHub