Projects
Personal Works
Projects, hackathon demos, and open-source.
Prompt Tuner
Cloud-based prompt optimizers send private data to external servers. Static libraries break with updates.
A Chrome extension that acts as a 'Grammarly for Prompts', optimizing chat inputs entirely on-device using Gemini Nano.
Impact: Built a fully offline optimizer that intercepts text directly from the browser DOM and tunes prompts dynamically.
Janus
Standard audio codecs fail below 6kbps, making satellite calls expensive and unintelligible in remote areas.
A 'Semantic Codec' that transmits meaning and tone rather than raw audio, using AI to rebuild voices over weak connections.
Impact: Achieved natural voice calls at just 300 bps, reducing satellite bandwidth costs while keeping vocal emotion intact.
CardinalCast
I wanted to build an end-to-end ML application to practice real-time data ingestion and job scheduling.
A weather prediction market where users place virtual bets on daily forecasts using ML-generated odds.
Impact: Built a daily NOAA data pipeline and a custom probability model that automatically resolves bets and updates wallets.
Prior Art
Standard LLMs often provide non-deterministic build-vs-borrow recommendations that lack empirical backing.
A build-vs-borrow intelligence system for determining whether to implement custom infrastructure or adopt open-source packages.
Impact: Provides deterministic package discovery and multi-dimensional health evaluation based on quantitative metrics and empirical research.
SchedSec
Pure LLMs often overlap tasks, ignore constraints, or violate dependency trees when scheduling.
An autonomous, serverless 'Scheduling Secretary' designed to automate the cognitive load of daily planning.
Impact: Uses a hybrid architecture: deterministic constraint solvers combined with LLM-based semantic placement.
OverSite
Traditional coding interviews usually ban AI. This fails to test how engineers actually work today.
An assessment platform that evaluates how developers use AI, tracking real-time IDE telemetry and ML models.
Impact: Built a scoring pipeline using XGBoost and Gemini to classify developer behavior and prompt quality from live IDE sessions.
Pulsera
Medical alert systems are purely reactive, notifying families only after an incident rather than preventing one.
A wearable and mobile health platform that detects physiological distress and intervenes autonomously using AI voice coaching.
Impact: Integrated WatchOS telemetry with a custom ML pipeline to detect distress and deploy real-time voice stabilization before any alert fires.
Contact
Feel Free to Reach Out
Always open to collab or discussing new opportunities. Check out my socials.