At A Glance
A focused tutor for the first signs that matter
Sanjog began as a Bachelor of Computer Science thesis. The team trained their own model to read hand signs from a webcam and kept building after college. The current scope is deliberate: alphabet and numbers, with instant feedback, so beginners can practice without a classroom.
This is not a story about adding more signs. It is a story about the first ten minutes of a learner’s life. If those minutes feel like a lecture, they leave. If they feel like a conversation with a patient tutor, they stay.
The product constraint was unusually sharp: a laptop, a browser, a webcam, and a model that only knows A–Z and 0–9. Everything in the interface had to make that constraint feel like a promise, not a limitation — “you can master this set” instead of “the product is unfinished.
Challenge
College Final-Year Project
Focus
AI-powered sign language learning
Real problem
No instant feedback for beginners
Goal
Learn, practice, and improve signs in real time
Deliverables
AI-powered web platform, custom AI model, browser-based recognition
The Challenge
How might we give beginners instant, encouraging feedback using only a webcam?
Most learning tools show a picture of a sign and hope the learner copies it. Sanjog had to close that loop in the browser — no kit, no studio, no stored video.
The design challenge was not “can the model detect a letter.” That was an engineering problem the team was already solving. The design challenge was: what happens in the second after the model answers? Shame, confusion, and dropout live in that second. Encouragement, correction, and the next attempt live there too.
We treated the webcam as a social object, not a camera. Asking for it is a trust event. Showing a live feed of someone’s hand is a vulnerability event. The UI had to earn both without a login wall or a privacy policy dump.
Constraint
Browser-only. No native app, no GPU assumption, no stored frames. If it does not run on a student laptop, it does not exist.
Job
A first-time user forms letter A, sees Correct, and knows why — inside one sitting, without a teacher in the room.
Failure
A wrong label feels like a personal failure. False negatives punish the exact people the product is for.
Role
Sequence the session, frame the camera, and make the model’s confidence readable as help, not a scoreboard.
Problem
The gap is not motivation. It is feedback.
Millions of deaf people use sign language as a primary language. Most hearing people never learn it. Books, videos and classes are costly, hard to reach, or passive — and none tell you if your hand shape is right.
Beginners need immediate feedback to know whether they are signing correctly.
Core design insight
Beginners do not fail because they do not care. They fail because they cannot see themselves. A YouTube video of an expert signing “A” does not tell you that your thumb is too high. A PDF chart does not tell you that you dropped the hand before the model or a teacher could read it.
That missing loop is how people decide “I am bad at this” after twenty minutes. The social cost is larger: family members who want to talk to a deaf relative, students in an accessibility course, volunteers who will never sit in a paid class. The product is a communication bridge, but the first job is much smaller make the first letter feel learnable.
Few free, interactive ways to learn
Most resources are passive videos
No instant feedback while practicing
Beginners find the first hours slow and uncertain
People
Who it is for
The product is built for first-time learners. If a family member, a student, or a teacher can succeed on a laptop in ten minutes, the model is doing its job.
We designed for one primary mental model: “I am going to look awkward on camera.” Everything else — XP, streaks, lesson structure — is downstream of making that feeling safe. Teachers and NGOs are welcome, but they are not the first session. The first session is someone alone at a desk.
Primary
Beginners learning basic signs, Students interested in accessibility, Family and friends of deaf people
Secondary
Teachers looking for a free practice tool, NGOs and the simply curious
Approach
Practice, then proof
The journey is short on purpose. Camera permission is explained, the first lesson is a single letter, and the loop of try → result → XP is visible within a minute.
We mapped the session as a story with a beginning, a beat of courage, a result, and a reason to come back. The beginning is permission. Courage is putting a hand in the frame. The result is Correct or Try Again — never a silent failure. The reason to return is an honest streak, not a dark pattern.
Scope was a product decision, not a compromise. Teaching the whole language on day one would have produced a dictionary. Teaching A–Z and 0–9 produced a skill. Recruiters often ask “why so little.” The answer is: a closed loop you can finish is more humane than an open world you cannot.
01
Discover
Free, in the browser
02
Allow camera
Privacy explained
03
Practice
Hold the sign
04
Feedback
Correct or try again
05
Progress
XP and streak
Open the site and allow the webcam
Pick Alphabet or Numbers
Show the sign to the camera
The model reads hand shape in real time
See Correct or Try Again with a confidence score
XP and streaks update
Research Insights: Insights sorted from research into observations, pain points and ideas.
User Flow: The path a user takes through the product, including the decisions along the way.
Solution
A product that looks like a tutor, not a demo
The interface is built around three surfaces: a calm homepage, a live recognition stage, and a progress view that makes improvement visible.
The homepage sells a feeling of lightness — start free, no card — because the camera ask comes next. The practice stage is the product: live feed, target letter, binary result, confidence as a percentage so “wrong” can still mean “close.” Progress is not a vanity dashboard; it is evidence that yesterday’s awkward A is today’s 92%.
Real-time AI recognition
Detects hand signs from the webcam with a custom-trained model running in the browser.
Alphabet and numbers
Structured lessons for A–Z and 0–9, with rising difficulty instead of an open dictionary.
Instant feedback
Correct or Try Again, plus a confidence score so the learner knows how close they were.
Gamified progress
XP, daily streaks and achievements turn a serious skill into a daily habit.
Privacy-first
Video is processed in the browser. Nothing is stored or sent to a server.
No download
Runs in the browser on a laptop or desktop. The barrier to the first sign is a URL.
Principles
These were not posters on a wall. They were rejection criteria. If a screen needed an account before a sign, it died. If feedback arrived after a quiz instead of during the gesture, it died.
Feedback in the moment
/
Learners should know right away if a sign is right. Delayed correction is how bad habits form.
Practice as a game
/
XP, streaks and achievements make return visits feel like a habit, not homework.
Private and frictionless
/
No download, no stored footage. If the camera is a worry, the product has already lost.
Make progress visible
/
Every practice session should show learners how far they have come.
Design for imperfect conditions
/
Feedback should remain helpful when lighting, hand position, or recognition accuracy varies.
Information architecture: How the product is organised, from entry points down to the deepest screens.
Decision
Each decision is a trade-off we can still defend. Confidence scores can intimidate; we kept them because “try again” without a number feels arbitrary. Streaks can guilt people; we kept them because skill learning without a cadence dies on day three. On-device inference is harder to ship; we kept it because a stored webcam feed would end the product in one news cycle.
Soft paper-adjacent surfaces, rounded cards, a restrained plum on product chrome, and a supportive tone. The goal is to make the first session feel light even when the skill is great.
Outcome
Sanjog is live. That matters more than a polished fiction. People can open a URL, allow a camera, and practice. The honesty of the outcome is that the vocabulary is still the first chapter of a language and that some team copy on the live site still reads as a student project. Both are true. Both are on the roadmap.
What works
Could improve
Next
What this project taught
A focused scope (A–Z + 0–9) is a smart way to validate an AI product.
Real-time feedback is a powerful differentiator in skill learning.
Gamification turns a serious topic into a daily habit.
Browser-based, privacy-first design removes the biggest barriers.
FAQ
Questions I get asked a lot
Anything else, just send me a note, I answer within a working day.