What I've built

Marks Automation for a Government Primary School

My father is a headmaster at a rural government primary school, and the marks for each subject were kept in separate notebooks. Entering them into the government portal meant searching each student across every register. I built a pipeline to remove that work.

• A browser extension reads the class metadata, and student roster from the portal.
• A server hosted on Render generates printable marks sheets per subject and reads the completed sheets back.
• Scanning extracts the handwritten values. Any value with low confidence is highlighted so the teacher can correct it manually.
• The extracted data is converted to JSON and filled into the portal automatically. I ran it against the real portal.

Teachers have been using it in a pilot. Their feedback so far is that they can write on the printed sheets without extra time, and the scanning removes the re-typing step. The product is meant to be free for government teachers, and I plan to keep maintaining it.

Echoloom: Dynamic RPG Game Framework

Requires an API key to generate content.

A 2D game built in Godot where the content is generated by an LLM at runtime: maps, NPCs, quests, and items. The main design choice is that the model outputs structured data (JSON), and the game engine deterministically parses and runs it. Creative generation stays separate from game logic.

• Map generation uses Perlin noise for surface biomes and a tunneling algorithm for underground corridors.
• Dungeon loot tiers come from a flood fill that ranks rooms by distance from the entrance.
• NPCs are generated in dependency order, so a quest's items exist before the quest references them. Dialogue trees change with quest state.
• A vector database stores past generated content and player decisions, so later story content stays consistent with earlier events.

The Flask backend uses LangChain, Groq, and HuggingFace embeddings. Enemy and sprite generation were left as future work, because I didn't find a free image generator that produced consistent animation frames.

3-DOF Robotic Arm

My first robot build. I 3D-printed the parts, powered the servos through a 3V AC-to-DC converter, and controlled everything from a Raspberry Pi 3B+.

• No existing library fit the build, so I derived the inverse kinematics myself using the DH parameter method, working through it in Excel.
• I wrote a local Python library that converts target coordinates into servo angles.
• Two control modes run from a laptop connected to the Pi: hand tracking with computer vision (index finger), and a gamepad joystick.

Dynamic Password Generator

A physical device, inspired by a scene in the Telugu film Nanna Ku Prematho. Pressing a button on an ESP32 generates a random one-time password, shows it on an LCD, and expires it after 10 seconds. The password is only visible on the device, so the physical device is part of the security model.

• Each password is single-use and time-limited.
• Current limitations, which I documented as future work: the device talks to the file-access system over plain HTTP (no encryption yet), and there is no lockout after repeated wrong passwords.

Zillow Home Value Prediction

A regression model that predicts home values. I followed a tutorial for the project structure, then made my own choices on feature selection (for example, that the surrounding area matters more than the color of a multi-floor apartment) and on the model. The project doesn't yet have a documented accuracy or error metric, which I plan to add.