Projects
A collection of systems I have built — ranging from real-time AI applications to full-stack web platforms. Each project reflects a focus on solving meaningful problems with clean, production-ready code.
6 projects
CrisisTriageAI
Real-time multimodal triage platform that analyzes text, voice, and live phone streams to detect emotional distress and classify risk levels. Runs entirely locally with no external API dependencies.
Demonstrates production-grade ML engineering with privacy-first architecture and real-time streaming constraints.
What I Built
- Designed modular FastAPI backend with WebSocket streaming for real-time audio and text processing
- Implemented local Whisper ASR transcription and Librosa-based prosody extraction (pitch, speech rate, pauses)
- Built custom DistilBERT-based neural classifier for risk-level classification with 73+ pytest tests
- Developed Next.js dashboard with live triage view, session history, and analytics aggregation
- Architected privacy-first system: all inference runs locally, ephemeral data handling, no cloud dependencies
Sub-second inference latency on local hardware; zero external API calls for core triage functionality
Tech Stack
ML & Public Health Modeling
Work in progress research project analyzing fairness gaps in public-health AI systems. Using CDC PLACES and U.S. Census socioeconomic data, I help build and evaluate baseline ML models that predict diabetes and cardiac-disease prevalence at the ZIP-code level.
Research into whether health prediction models systematically underpredict risk in low-income communities.
What I Built
- Building baseline ML models for diabetes and cardiac-disease prevalence prediction at ZIP-code level
- Analyzing CDC PLACES public health data combined with U.S. Census socioeconomic indicators
- Evaluating model fairness across income brackets and demographic groups
- Investigating systematic underprediction patterns in low-income communities
Tech Stack
Facade Risk Analyzer
AI-powered building assessment platform that processes facade images to detect structural defects, assign risk scores, and generate condition reports with repair cost estimates.
End-to-end AI application with pluggable analyzers, automated reporting, and production-ready CI/CD.
What I Built
- Designed modular FastAPI pipeline with pluggable analyzers (OpenAI Vision, Mock, Replay) for flexible deployment
- Implemented SQLite-backed jobs engine with output versioning and automated PDF report generation
- Built Next.js frontend with job dashboards, analysis history, and shareable report links
- Created CI/CD workflows with GitHub Actions, reproducible fixtures, and CLI tools for local orchestration
- Delivered working platform that processes images end-to-end from upload to downloadable PDF report
Complete image-to-PDF pipeline in under 30 seconds; deterministic replay mode for testing and demos
Tech Stack
ERP Data Extraction Tool
Desktop automation tool that extracts bank balances from ERP software and exports to Excel on a configurable schedule. Built during internship at Edenred Turkey.
Real-world automation solving an actual business problem with measurable time savings.
What I Built
- Developed Python automation using pywinauto to extract data from legacy ERP UI
- Built PyQt5 interface for scheduling, file management, and configuration
- Implemented robust logging and error handling for unattended long-term operation
- Delivered working tool adopted by treasury team to eliminate daily manual reporting
Eliminated 30+ minutes of daily manual work; reduced data entry errors to near-zero
Tech Stack
Lumen Search and Spaces
AI-powered search and research platform combining multi-model inference with web intelligence. Generates sourced answers and structured long-form reports.
Production multi-model orchestration with real-time web retrieval and citation generation.
What I Built
- Integrated model-routing logic across DeepSeek-R1, LLaMA-3.3-70B, and O3-Mini for query-specific optimization
- Implemented search tool integrations with Exa and Firecrawl for document and web retrieval
- Built Lumen Spaces feature for autonomous research with structured report generation
- Contributed to multi-model orchestration pipeline that produces sourced, citation-backed answers
Reduced hallucination rate through grounded retrieval; improved answer quality with model-specific routing
Tech Stack
Blaze AI Shopping Assistant
Voice-enabled AI shopping assistant that extracts product requirements from natural conversation, performs real-time web searches, and provides personalized recommendations.
Conversational AI with persistent memory, real-time search, and multi-turn reasoning.
What I Built
- Implemented LLaMA-based reasoning for product requirement extraction from voice queries
- Built memory layer to track user preferences and purchase history across sessions
- Integrated real-time web search for product discovery and price comparison
- Developed on-the-fly product analysis to generate personalized recommendations
Context-aware recommendations across sessions; natural voice interaction with sub-2s response time