← Selected work

01 / 04Computer vision · Real-time

Real-time Sign Language Translation

[One sentence: what it does and for whom, e.g. Live camera-to-text translation of ASL fingerspelling and common signs at 30 fps on a laptop.]

Year
[2025]
Role
[Lead developer · vision pipeline, UI]
Team
[Solo / 3 people]
Tools
[Python, PyTorch, MediaPipe, OpenCV]
Links
GitHubDemo video
Still from the live translation demo
Still · live demo (1600×900)replace
Fig. 0[Caption: what the viewer is looking at, e.g. the live view with predicted gloss and confidence.]
01Context

[2–3 sentences of context: who signs, why real-time matters, what existed before and where it fell short.]

[One sentence on the scope you set: vocabulary size, latency budget, hardware target.]

02What I built
  1. [Built the capture → landmark → sequence model pipeline end to end]
  2. [Designed the temporal model / chose the architecture and why]
  3. [Collected or curated the dataset: N signers, N clips]
  4. [Built the demo UI and ran the user test]
03How it works

[Paragraph: the key technical idea, in plain words first, then the specifics. Name the trade-off you made for latency.]

[Paragraph: how you validated it as you went.]

Figure · pipeline diagramreplace
Fig. 1[Pipeline diagram: camera → hand/pose landmarks → temporal model → gloss → text.]
Figure · resultsreplace
Fig. 2[Confusion matrix or latency breakdown.]
04Outcome
top-1 accuracy on [N] signs[92%]
end-to-end latency[45 ms]
signers in the test set[12]

[What the numbers mean and where it still fails. Honest limits read well.]

[Optional one-line pull quote from a user or from you.]
05Reflection

[Two or three sentences: what you would do differently, what surprised you, what this made you want to study next.]

GitHub ↗ Demo video ↗