Surya Kasturi
Writing on human condition
Healthcare
- Healthcare's infrastructure gets all the attention, but the patient experience stays broken. The cost of building software has dropped enough that it finally makes sense to just solve one obvious problem for one person at a time without overthinking market size. That's what I'm building toward. Free and open source consumer healthcare products.
- Florina, 2025 - present
Private and voice-first period tracker app for women across life stages. The data stays on the device, with minimal design, to track a wide spectrum of symptoms using the OpenAI Whisper and Google Gemma models. The current apps are too cumbersome, not designed to help people with chronic conditions, going through transitory phases, have significant privacy issues, and are costly. This product is positioned to differentiate from this paradigm. Currently available for beta-testing. Please download here to try and share your feedback.
Deep learning
- How Far Can One Example Go?
Reinforcement Learning Across Logic, Math, and Language, Sundai Club, Jun 2025
Can reinforcement
learning on a single high-variance summarization example improve performance on other, unseen instances within the
same domain?
- Schema guided dialog state tracking, DSTC8, AAAI
2020
Developed dialogue state tracking models suitable for large-scale virtual assistants, with a focus on
data-efficient joint modeling across domains and zero-shot generalization to new APIs.
- Synthetic dataset for document visual question
answering, Apr 2020
Developed a synthetic Visual Question Answering (VQA) dataset from existing
Question Answering (QA) datasets to evaluate the cross-modality generalization capabilities of models.
- Machine Reading Comprehension,
Dec 2019
Optimized Machine Reading Comprehension models on SQuAD, CoQA, and ShARC by analyzing and mitigating
the impact of entity-type sampling bias leading to significant performance gains. Top results from PAII Labs/Gamma
Lab on SQuAD.
- Knowledge-grounded conversation
modeling, DSTC7, AAAI 2019
Trained BERT and Memory Networks to generate contextually rich responses by
conditioning on conversation history and Wikipedia facts, with performance significantly boosted by normalizing
special characters in the Reddit dataset. Top results in the competition.
- Synthetic generation of MNIST style
dataset using Google Fonts, Jun 2018
Developed a synthetic MNIST style dataset using Google Fonts to
measure out of distribution performance of convolutional neural nets
- Knowledge distillation of time series forecast models,
Aug 2017
Developed deep neural net models for forecasting and detecting anomalies in chiller plants
performance using real-time sensors
Software tools
Open source
Links