Janhavi Dadhania

ML Engineer @ project44

building models that move things through the real world. writing about the systems underneath everything else.

Janhavi Dadhania

# about

I spend my days getting transformer models to predict when trucks will show up (harder than it sounds) and my evenings wondering whether consciousness is computable.

Before logistics I was at Razorpay routing payments with reinforcement learning, and at ShareChat figuring out face detection on phones that cost less than a textbook. The through-line is: I like problems where the real world refuses to cooperate with your model.

Outside of work I read too much (economics, philosophy, history, whatever's lying around), run two Substacks about complex systems and the strange loops in everything, and occasionally get lost in papers about foundational models and multimodal architectures.

IIT Dharwad CS, class of 2022. Arctic Code Vault Contributor on GitHub (which mostly means I was in the right repo at the right time, but I'll take it).

  location   : bangalore
  focus      : deep learning, RL
  reading    : too many books
  writing    : two substacks
  github     : arctic code vault
  cgpa       : 8.7/10

# work

2025 - present

project44 ML Engineer 2

Building a transformer-based ETA prediction model for logistics trucks. Improved accuracy with lane embedding and hex-level geospatial features. Also built an NL2SQL chatbot using LLM-based RAG that automated customer data queries, cutting latency from 30s to 20s.

2023 - 2025

Razorpay Data Scientist

Designed RL-based payment routing algorithms handling hundreds of transactions per second. Pushed success rates from 80% to 93%. Built an XGBoost model for predicting order returns before they happened.

2022 - 2023

ShareChat & Moj ML Engineer

Optimized deep face landmark detection for low-end mobile devices. Built image-text search across a 3M item fashion catalog for video commerce.

internships

Entrupy + Ultrainstinct

Siamese networks for 3D point cloud registration at Entrupy. Multimodal deep networks for video captioning on ActivityNet at Ultrainstinct.

# research

arXiv:2112.13706

Multi-Image Visual Question Answering

BTech project at IIT Dharwad. Proposed a new dataset for VQA with multiple image inputs and a single ground truth. The final model used ResNet + RCNN image features with BERT embeddings, inspired by stacked attention networks. Achieved 39% word accuracy and 99% image accuracy on CLEVER+TinyImagenet.

arXiv:2101.05806

Video Captioning on ActivityNet

Introduced WAFTM (Weighted Additive Fusion Transformer with Memory Augmented Encoders) for video captioning. Used a novel feature fusion method and memory-augmented transformer encoders. CIDEr of 92.4 on MSVD, METEOR of 0.091 on ActivityNet.

# writing

Complex Systems in the World

The more serious one. Chaos theory, consciousness, Hindu cosmology meets quantum mechanics, how minds simulate reality. Basically: what are the rules underneath the rules?

Personal Substack

The everything-else one. Language preservation through a math lens, AI personified as narrative, culture as organism, old scientific diagrams that happen to be beautiful.

# reading

A partial list. The full one would crash your browser.

economics & systems

india & politics

philosophy & everything else

  • The Story of Philosophy — Will Durant
  • No Logo — Naomi Klein
  • Animals & Men — David Katz
  • The Extreme Future — James Canton
  • Outline History of English Language — Frederick T. Wood
  • Play Bigger — Ramadan, Lochhead, Peterson, Maney
  • Getting Things Done — David Allen

# contact

  $ cat contact.txt

  email    : janhavivdadhania@gmail.com
  linkedin : janhavi-dadhania
  github   : JanhaviDadhania
  substack : complex systems | personal