01
Normalized Evaluation
Metrics must be comparable. I build normalization layers to standardize disparate scores into unified scales for clear decisions.
Work
Selected projects that reflect my focus: clear metrics, reliable pipelines, and actionable AI insights.
PrepGenie: LLM Interview Coach
Edge-native interview platform leveraging Pyodide/WASM and recursive AI failover. Delivers resume-aware RAG and real-time LLM-as-Judge scoring for contextual mock interviews.
120msAvg TTFB
<100msCold Starts
99.9%Uptime
3-ModelFailover Chain
Pyodide/WASMEdge NativepypdfLLM-as-JudgeRecursive Failover
Technical Challenge
Ensuring zero-downtime coaching sessions. I engineered a Gemini to Groq to Mistral recursive failover chain with async request handling and distributed SQL to guarantee reliable performance.
LLM Content Evaluation Framework
Unified DeepEval pipeline benchmarking GPT-4, Gemini 2.5, and LLaMA 3-70B across 7 normalized metrics for objective, data-driven model routing.
7Metrics Normalized
96%GPT-4 Accuracy
40xGemini Cost-Efficiency
DynamicModel Routing
DeepEvalPythonBLEU/BERTScore/GEvalCustom Normalization
Production Insight
Revealed domain-specific routing strengths by benchmarking across 20+ real-world articles, enabling a production-ready pipeline that balances precision, cost, and creative output.
Autonomous AI Penetration Testing
Multi-agent cybersecurity framework orchestrating Reconnaissance, Vulnerability Assessment, Exploitation, and Strategic Reporting using MCP and Llama 3.3.
95%SQLi Detection
90%XSS Detection
$10K+/yrScanner Cost Saved
4 AgentsOrchestration
MCPLlama 3.3 70BCloudflare TunnelsOWASPDocker
Technical Challenge
Architected a hybrid cloud-edge pipeline that integrates industrial tools like Nmap and sqlmap via secure tunnels while enforcing strict ethical guardrails against internal IP pivoting and restricted domains.
Real-Time Stocks Serverless Dashboard
Global serverless financial dashboard utilizing V8 Isolate runtimes and React 18. Features on-demand D1 SQLite caching to prevent upstream API exhaustion.
<60sData Latency
90%API Call Reduction
ZeroIdle Compute Cost
D3.jsSVG Charting
React 18 + ViteCloudflare D1 + R2Finnhub APIOn-Demand Caching
Technical Challenge
Designed a pull-model authorization engine that intercepts requests via MAX timestamp queries. Resolves 9,999 subsequent concurrent reads entirely from edge cache within each 60-second window.
GestureSynth: Real-time Hand Gesture to MIDI
Vision-based system using Python and MediaPipe that translates hand gestures into MIDI output with low latency. Features LLM-powered adaptive tuning for optimal playability.
30 FPSPerformance
<80msLatency
40%Faster Learning
LLM-AdaptiveTuning
PythonMediaPipeOpenCVMIDIPyQt5
Technical Challenge
Calibrating gesture-to-music mapping in real-time. I built an LLM-powered adaptive tuning module that analyzes note stability metrics to automatically optimize parameters for improved playability.
Manifesto Explainer: Political Party Analysis
NLP platform transforming dense political manifestos into accessible insights using sentiment analysis, word clouds, and Llama-3 for contextual understanding and auto-translation.
10xFaster Scanning
ContextualUnderstanding
MultilingualAuto-Translation
DeployedWeb App
Spring BootPythonNLTKLlama-3Heroku
Technical Challenge
Manifestos are messy. I integrated Groq's Llama3-8b model to summarize what parties actually promise, turning static documents into interactive policy assistants with symbol detection.