SoumyaPatel

Syllabus parsing requires extracting highly structured timelines from messy PDFs without blowing up API token costs, while handling real-time voice inputs on the frontend and keeping strict academic integrity safeguards.

Built a Next.js frontend with Tailwind CSS and Framer Motion, integrated with a FastAPI Python backend. Used the browser's Web Speech API to build a voice onboarding flow, set up AWS Cognito for auth, uploaded documents to S3, and called Gemini AI. Made custom prompt guardrails and a hybrid caching/routing engine to reduce API usage by 70%.

Deployed a functional web app that helps students onboard in seconds using voice commands. Integrated speech recognition APIs, optimized LLM costs, built prompt guardrails, and scaled cloud infrastructure.

Natural language combinatorics questions exhibit highly varied grammar, making traditional rule-based parsers fragile. Conversely, utilizing heavy LLMs introduces unwanted latency, API costs, and overhead for extracting simple algebraic parameters.

Built a Flask backend powered by scikit-learn. Developed a custom NLP pipeline using an SVC to classify the problem type, and combined it with a Regex parser to extract numerical constants and output the corresponding LaTeX math formulas.

Reached 92% classification accuracy to provide instant math solutions. Integrated scikit-learn models with a Flask backend and rendered LaTeX math formulas dynamically on the frontend.

Manual client mod management is error-prone and version-dependent, while unoptimized Java settings cause periodic in-game lag spikes (garbage collection stutter).

Built a desktop client using Electron and Node.js that communicates with the Modrinth API to resolve and install Fabric mod JARs directly to the user's local directory via Node's fs and https modules. Validates ZIP headers and EOCD signatures at the binary level to detect and auto-repair corrupted downloads. Automatically applies low-latency GC flags (UseG1GC) and caps frame rates to eliminate micro-stuttering. Integrated minecraft-launcher-core to handle authenticated game startup, and set up Electron IPC channels via a preload script to stream real-time launch status back to the renderer UI.

Automated local game environment setup. Managed third-party APIs (Fabric/Modrinth), handled file downloads, persisted local configurations, preventing corruption crashes and tune JVM garbage collectionand worked with Electron's main/renderer IPC bridge.

Writing database scripts, its easy to make bugs like leaving connections open, letting user input break the queries, or having a script fail halfway and leave the data in a mess.

I separated the database code from the main program logic. All database setup and connections are handled in a single class (database.py), while the main script (main.py) only needs to call simple functions like fetch_all(), execute_query().

I learned how to use Python's context manager (with statements) to automatically close database connections and how to use query parameters to keep inputs safe.