HackStack

Project Info

TCM Coach

Devpost

Interactive guidance for TCM self-care routines.

Inspiration

Interactive TCM Coach was inspired by a real self-care problem from our team: someone wanted to try hand acupressure for period discomfort, but did not know whether the position, rhythm, or movement was correct. Static diagrams and videos were not enough. We wanted to build a system that gives immediate visual feedback while keeping clear safety boundaries. What It Does Interactive TCM Coach uses AR to guide users toward the correct hand acupressure area, then uses computer vision to analyze pressing rhythm, accuracy, and stability in real time. Instead of only showing where to press, it helps users understand whether they are doing the routine correctly. The demo also includes a front-end 3D body model with meridian visualization, so users can see the broader route/context of a point instead of only seeing a flat instruction. How We Built It We built a web-based demo with React, camera input, MediaPipe hand tracking, AR-style overlays, front-end 3D modeling, and computer vision logic for hand position and pressing frequency. We also created a video-to-JSON pipeline to extract fingertip position and rhythm features from training videos, which can support future model training. Challenges We Ran Into Fine-tuning a CV model to recognize exact acupressure position was too large for the hackathon scope. Our workaround was to use AR target circles to mark the intended area, so the software does not need to fully identify the acupoint from scratch. It only needs to check whether the user's finger reaches the marked region. Overlapping fingers made fingertip-level tracking noisy. Frequency and stability detection were also inconsistent: sometimes the signal worked, but sometimes it broke when the hand moved quickly or the finger was partially hidden. Since AR handles position guidance, we simplified frequency detection by measuring motion around the target region, such as centroid motion frequency, instead of requiring perfect fingertip landmark detection every frame. Fitting long meridian paths onto a 3D human model was difficult because the line has to follow the body surface instead of floating through the model. Our workaround was to split each long meridian into smaller segments, then fit and adjust the segments one by one. iOS camera access was also tricky during development because Safari is strict about camera permissions and secure origins. We had to adjust the local setup so the phone could open the demo with camera access instead of only testing on desktop. Accomplishments That We're Proud Of We are proud that we kept the idea small enough to actually demo. Instead of trying to build a huge health AI product, we focused on one simple user pain point: people can read acupressure instructions online, but they still do not know if they are pressing in the right area. We are also proud of the workaround we found for the vision problem. Exact acupoint detection is hard, especially in a hackathon, so we used hand landmarks and rough target regions to make the guidance understandable enough for a working demo. What We Learned We learned a lot about a new kind of software engineering workflow with AI. The challenge is not just writing code faster, but learning how to collaborate, define success, keep the right context, and design useful evaluation systems while the project is moving quickly. We also learned how to get the key information needed to contribute even when we did not fully understand every part of the project. Instead of trying to know everything first, we focused on the important constraints, acceptance criteria, and failure cases, then used those to guide the build. What's Next for Interactive TCM Coach Session: Expand from one-time pressing feedback into guided course-style sessions, where users complete structured routines over time instead of a single action. LLM: Add an assistant with RAG over curated TCM and safety content, so users can ask self-care questions while the system stays grounded in reviewed sources. More Content: Support more acupressure points, routines, and hand/wrist regions, with AR guidance, computer vision feedback, and progress summaries across sessions.

Analysis

Compare with all teams

View

Metric

Figures cover GitHub contributors during the hackathon window. A co-authored commit counts in full for each author, so per-member totals add up to more than the whole-team figures.

Technology

Found in codeClaimed only
  • CSSIn code
  • HTMLIn code
  • JavaScriptIn code
  • PythonIn code
  • ReactIn code
  • SwiftIn code
  • Tailwind CSSIn code
  • TypeScriptIn code

8 of 8 appear in the indexed code.

AI coding agents

  • Claude CodeConfig · Commits

Detected from committed agent config files and commit authorship. Absence of a signal is not proof an agent was unused.

Codebase size

Source size

491 KB

Source files

120

Counts recognized source files only; vendored directories, binaries and lockfiles are excluded, so this is smaller than the repository on disk.

0 stars