Building MissionX at Junkbot Robotics
I work on MissionX, a learning platform for school students. I build its career simulations where children do a real job for a week, learning games that teach AI up to grade 12, our own block-coding editor for micro:bit, the AI agent that helps students and teachers, and the analytics that show where they get stuck.
- Role
- AI Software Developer
- Since
- June 2025, 1 year 5 months
- Where
- Full-time, hybrid, Kasaragod
- Users
- School students aged 7 to 18, and their teachers
Career simulations
Future Jobs. Sign in at MissionX to play.
Each simulation drops a child into a real job for a week. The rule I design by is do the job, don’t answer questions about it. Nobody is asked “what is overfitting?”. They press Train round after round and watch the quiz score bend down while a parrot squawks “I memorised it!”.
Every career has four levels on a world map, a buddy character with its own bar at the bottom of the screen, and a finale. Where I can, the finale is a real incident retold for a ten-year-old.
Cybersecurity Analyst
Defend Nova Academy across a campus map.
Sort real emails from fakes, decide which logins to allow, fight a virus in the computer lab.
Finale. A boss fight in the Main Hall.
Embedded Systems Engineer
Build BOLT, a test rover, at Sparkworks Lab.
Choose parts on a budget and wire them, snap code blocks into a loop, probe noisy wires and plan a battery.
Finale. Based on NASA’s 1997 Mars Pathfinder fix: find why a rover keeps restarting and send a checked update across space.
IoT Specialist
Connect the city of Lumen Bay before its harbour festival.
Pick and place sensors, give each device the right network, cover the streets with as few gateways as possible.
Finale. Based on the 2016 Mirai botnet: quarantine hijacked cameras and roll out a security update in safe waves.
Quantum Computing Engineer
Get a quantum lab’s machine working; the level map is the fridge itself, colder the deeper you go.
Shield qubits from noise, calibrate pulses on a Bloch sphere, entangle two qubits with gates.
Finale. Read parity alarms and fix flipped qubits as the error-correcting code grows from 3 to 5 qubits.
Machine Learning Engineer
Train ZURI, a camera-trap AI in a wildlife reserve.
Label photos, hide a quiz set, stop training before it memorises, find the shortcut it learned.
Finale. Data drift arrives as a night fog. Retrain, set the alarm line, and leave the final call to a person.
Two ideas inside every simulation
Rebuilt here from scratch so you can try them. Not the platform’s code.
A help ladder. Being stuck should never end the game, and asking for help should never feel like failing. Hints come in three rungs, a nudge, a point and a clue, and they are free. Only taking the answer costs something: it caps the mission at two stars.
A mentor that forgives spelling. Children type “fishing” when they mean phishing and “malwear” for malware. The mentor matches words by how they sound and by edit distance, entirely in the browser, so nothing a child types leaves the device.
Mission: is this email real?
From: Nova Academy IT <help@nova-acadmey.co>
Subject: URGENT: your account closes today
Click here within 1 hour and type your password to keep your account.
Stuck? Hints are free. Only the answer costs a star.
AI Textbook games
Short missions that teach AI ideas through play, for school students up to grade 12. The early grades sort things and follow simple rules; the later ones build neural networks and fine-tune models.
Each game is a small web app with no game engine, and games never share code: a new one starts as a copy of the closest existing game. I wrote about why in How I build learning games without a game engine.
Junkbot Visual Code
code.junkbot.co, public
Students program their micro:bit robots with blocks. Instead of sending them to Microsoft’s editor, we run our own: I built Junkbot Visual Code on Microsoft MakeCode for micro:bit, which is open source (the pxt-microbit target on top of the PXT framework).
The hard part was adding Junkbot sign-in and saving projects to our own backend, inside a codebase far too big to hold in your head. I was stuck for 25 days. What fixed it was writing down how MakeCode already handles sign-in and storage before changing anything; after that it took hours. The story is in The context problem.
- Based on
- Microsoft MakeCode for micro:bit (pxt-microbit, PXT)
- Added
- Junkbot sign-in, and projects saved to our own backend
- Branded
- Junkbot name and look, served at our own address
- Used by
- Students programming micro:bit robots in class
The conversational AI agent
An assistant that helps students and teachers with robotics projects. Shipping it taught me that the model is the easy part. What broke in production was everything around it:
- Latency. Streamed replies, fewer API calls, and requests run in parallel where they can be.
- Invented components. Structured retrieval, so it stops confidently recommending parts that aren’t real.
- One agent doing everything. Split into workflows with clear responsibilities.
- Slow reads. A thread-safe in-memory cache took user lookups from 100–300 ms to under 1 ms.
- Checkpoint bloat. A scheduled cleanup for old conversation checkpoints that were slowing everything down.
The full story, problem by problem: What shipping an AI assistant for STEM education taught me.
Analytics that respect children
I integrated PostHog so the team can see where students stall, which hints they need and when they retry, without collecting anything about who they are. The players are children, so the defaults are strict:
- Autocapture and session recording stay off.
- Only a fixed list of events can be sent, and the types won’t compile if a new field sneaks in.
- When a child asks the mentor a question, only the kind of match is sent, never the words.
- Mission started
- which mission, which mode
- Action taken
- right or wrong, and time since the last action, which shows stall points
- Hint used
- which rung of the ladder
- Mission failed
- where, and whether they try again
- Mission complete
- stars, time taken
What I use there
- Games and simulations
- React, TypeScript, Vite, Vitest, PostHog
- AI agent
- Python, FastAPI, LangChain, LangGraph, LangSmith, Next.js, Supabase, AWS
- How a game starts
- A design document first: learning goals, then levels, then code
- Testing
- Game rules as plain functions with their own tests