Exploring Technology Through Real-World Problems
Led by Ansh Gupta, Founder of Together With Us, the Innovation Lab showcases independent research and technology projects exploring challenges in healthcare, public health, accessibility, and student well-being.
These projects reflect Ansh's curiosity about how technology can address real-world problems, and the importance of learning through experimentation, testing, and feedback. They range from research studies and working prototypes to early-stage ideas that continue to evolve.
Project 1. Refluent — AI-Assisted Communication for People with Aphasia
Refluent is an experimental AI communication tool developed by Ansh to help stroke survivors with non-fluent aphasia express themselves more easily. Using a fine-tuned transformer-based language model, it explores how fragmented speech can be reconstructed into more complete sentences while preserving the speaker's intended meaning.
Pilot Findings: What We Learned
Refluent was evaluated with speech-language pathologists in India and the U.S., each with 30+ years of clinical experience, including clinicians at Bhutani Speech and Hearing Aid Center, New Delhi, India, and SPARC Health, Seattle, USA, using examples drawn from real aphasia cases.
The pilot challenged one of our original assumptions: AI-based speech completion is unlikely to help many people with moderate-to-severe aphasia because they may be unable to formulate enough of an intended sentence for a model to reliably infer meaning. When speech is highly fragmented, the model may generate incorrect completions rather than recover the speaker’s intended message.
Clinicians saw greater potential for people who can already produce most of an intended sentence—roughly 70–80%—but struggle with fluency or completion. In these cases, AI may be better positioned as an assistive communication aid, rather than a replacement for speech-language therapy.
The pilot also showed that the usefulness of an AI communication tool depends heavily on the patient’s social and cultural environment. In India, clinician feedback suggested that strong family involvement, greater availability of caregivers, and a preference for human interaction may create a meaningful adoption barrier for device-based assistance in some settings. In contrast, clinicians saw greater potential in contexts where patients are expected to communicate more independently, or where access to speech-language pathologists is limited. A U.S.-based speech-language pathologist specifically identified rural and underserved communities as a possible use case where an assistive tool could supplement—not replace—professional care.
These findings are shaping Refluent's next iteration, with a narrower focus on patients who may benefit most from AI-assisted communication.
Project 2. Overdose Hotspot Prediction — Machine Learning for Public Health
Ansh developed a spatial machine learning framework to identify geographic patterns in drug overdose risk and explore how predictive models could inform public health resource allocation.
Using public health and demographic data for King County, Washington, he evaluated XGBoost and Tweedie GLM models and developed a Budget Prioritization Rate (BPR) to connect predicted risk with potential resource allocation decisions.
He also evaluated the framework using Los Angeles data to explore its transferability across geographic settings.
- NeurIPS 2026 Trustworthy AI for Good Workshop, Paris : Invited Presenter, December 2026
- Washington State Public Health Association Conference : Accepted Presenter and Full Scholarship Recipient, October 2026 | Poster
- Washington State and Central Sound Regional Science Fairs: Second Place, Social & Behavioral Sciences, 2026
Project 3. ASL Avatar — Exploring Speech-to-Sign Translation
Ansh explored the use of machine learning to translate spoken language into digital representations of American Sign Language, developing an initial LSTM-based approach.
Recognizing the complexity of representing hand movements, facial expressions, and linguistic nuance, he subsequently developed a 10-page technical proposal outlining an enhanced FSQ-based representational learning approach and submitted it to the MIT THINK program.
Project 4. Safe Circle — Anonymous Support for Students
Safe Circle is a web-based platform developed by Ansh to provide students with an anonymous space to share academic, social, and personal concerns, including bullying.
The project explores how AI-assisted categorization and pattern analysis could help schools identify recurring concerns and strengthen student support systems, while recognizing the importance of privacy, confidentiality, and responsible intervention.