Neel Kirpalani
I'm a senior at Hong Kong International School. I build systems that take a messy signal from the real world (the motion of a skateboard, a face, an object on a table) and turn it into something a person can actually act on.
Most recently I founded REMAKE, an AI beauty app that's live on the App Store with real users and revenue.
I captain 936A, our VEX robotics team. We took an Excellence Award, placed third in Hong Kong, and qualified for the World Championship.
Before that I built EcoWise, an app that rates the environmental impact of everyday objects, and won Best Overall at a school hackathon with a digital lost & found.
Away from a keyboard I DJ as DJNK_HK, play guitar, and play tennis.
Projects
Point a phone at your face and it tells you which foundation shade actually matches, whether a product is counterfeit, and what is in it that might react with your skin.
- Founder & Co-CEO. Product concept, model selection, and taking it to the App Store.
- Mobile · Computer Vision
- Live on the App Store with active users and revenue
The problem
Buying foundation online is guesswork. Shade matching is unreliable, counterfeit beauty products are common, and the ingredients that cause reactions are buried in unreadable labels.
The approach
One camera input, three answers: a shade match from a face scan, a counterfeit check, and an allergen read on the ingredient list. Being right mattered more than being clever, so the model work went into accuracy on real faces and real product photos.
The outcome
Shipped to the App Store and in use by real people, with revenue. It is the only project here with users who are not me.
A sensor package on a skateboard deck that measures speed, landing force and ollie height in real time, and recognises the trick you just landed.
- Sole developer. Hardware, firmware, and the trick-recognition logic.
- ESP32 · IMU · GPS · BLE
The problem
Skateboarders have no feedback loop. You land a trick badly and the only signal is your ankles. Progression is guesswork and injury risk is invisible until it is not.
The approach
An ESP32 with an MPU6050 IMU and a GPS module is mounted to the deck. Orientation, speed and acceleration stream over Bluetooth Low Energy to a phone for live feedback. The same acceleration and gyroscope traces are pattern-matched to detect specific moves, so a kickflip is recognised from its signature in the data.
The outcome
Working hardware and a functioning data pipeline running on an actual board, producing board orientation, speed, landing force, jump height, and trick detection.
What actually happens to employment when robots and AI arrive, and what a workforce needs in order to adapt instead of being displaced.
- Independent researcher and author.
- Research · Labour economics
- Published in Scholarly Review Journal
The problem
The public argument about automation and jobs is conducted almost entirely through anecdote. The empirical literature is more specific and considerably less dramatic.
The approach
Synthesised findings from 20 empirical economics and labour-market studies, including Filippi et al. in Technological Forecasting & Social Change and Kolev et al. in the Proceedings of the Technical University of Sofia, to identify where displacement risk concentrates by sector and which reskilling strategies respond to it.
The outcome
Published as 'Robots at Work: A Comparative Analysis of Job Displacement in the Automobile and Healthcare Sectors' in Scholarly Review Journal, Fall 2026, Issue 17.
Scan an object and it tells you how bad it is for the planet, and exactly how to dispose of it.
- Sole developer.
- Mobile · MobileNetV2
- Roughly 80–90% classification accuracy
The problem
Knowing that you should recycle is useless if you do not know which bin a specific object belongs in, and generic advice does not survive contact with the actual object in your hand.
The approach
MobileNetV2 was the right tool here. Inverted residuals and linear bottlenecks keep the model fast and small enough to run on a phone, which is where this app is useful. The classifier output feeds a recycling guide that returns instructions for the specific item.
The outcome
A working mobile app that rates an object's environmental impact and returns disposal instructions, at roughly 80–90% accuracy.
A lost-and-found system for a whole school: mobile app, web interface, and one central database, so an item that gets handed in actually reaches its owner.
- Product concept, UI/UX design, and system workflow planning.
- Mobile · Web · Database
- Best Overall, HKIS SDLT Hackathon
The problem
Lost property at a school of this size is a physical pile with no index. Items are handed in and never matched to the people looking for them.
The approach
A team build combining a mobile app for reporting, a web interface for staff, and a central database so both sides are querying the same records. The workflow made it win.
The outcome
Won Best Overall across a school-wide hackathon.
Predict sale prices from 79 features of the Ames housing dataset, and get as close to the bottom of the leaderboard as possible.
- Solo competitor.
- Python · scikit-learn
- Top 2.6%, 113th of 4,382
The approach
Built composite features (total living area, age and quality indices), encoded categorical attributes such as neighbourhood, and cleaned missing and outlier values before training regularised linear and gradient boosting models. Optimised against RMSE between predicted and actual sale price.
The outcome
113th of 4,382 entrants, the top 2.6%.
Awards
Placed 3rd in Hong Kong and qualified for the VEX World Championship.
Won for a digital lost & found system built across app, web and database.
113th of 4,382 entrants.
Robots at Work, a comparative analysis of job displacement in the automobile and healthcare sectors.
- AP Scholar with Honor
Grades 9, 10, 11.
Experience
Rebuilt the corporate website for a Shenzhen smartwatch manufacturer and built product pages for seven models. Time on the factory floor doing assembly and quality-control testing shaped how the products were described.
One of seven students selected. Built automated workflows to replace manual data entry in the intern application and content submission pipelines, then used social media analytics on click-through and event participation to redesign website pages and content.
Two weeks on core computer science, data analysis and the mathematics behind machine learning. Worked with supervised learning models including linear regression and basic neural networks, applied to image, speech and real-world data, then presented to 30+ peers and professors.
Covered recent progress in AI and machine learning, real-world applications, and the ethical and societal implications of advanced AI systems, through group programming and presentation-based projects.
Activities
Promoted from Team Leader-in-Training. Runs robot design, programming and match strategy for eight people, manages build and testing timelines, and mentors junior members.
Taught LEGO-based robotics to underprivileged primary students, leading build challenges built around confidence and teamwork.
Over 100 hours of kennel care, feeding and socialising dogs to improve adoption readiness, plus weekend adoption drives and talking to prospective adopters.
Over 100 hours promoting animal welfare and plant-based living, designing the posters and social graphics behind it.
Performed at more than 30 events including Dragon-i and school proms. Live mixing across genres.
Twelve years of guitar. Selected for the senior performance team inside a 100+ member programme, with roughly seven live shows.
Competing in the Hong Kong Men's B Division, training several times a week on singles and doubles strategy.
Education
- Hong Kong International School
