I build AI-powered
systems & products

NesterAI Conversational System: Architected emotion-aware voice AI from scratch, achieving 90% cost reduction ($0.33→$0.01/min) with sub-200ms latency

Acoustic Pipeline: Benchmarked Deepgram Nova-2, OpenAI Whisper, Google/AWS STT. Selected Deepgram for sub-300ms latency with 8.4% WER at 2x performance-to-cost ratio

Hybrid Emotion Detection: Built Emotion Engine combining wav2vec2-MSP-PODCAST with LLM text sentiment, capturing 100% of communication vs. industry 50%

Custom Framework Extensions: Engineered EmotionForkProcessor, ContextInjector, and ToneModulationTTS as custom Pipecat processors

Hysteresis Controller: Designed stability controller with 2-frame lock, 1000ms rolling window, and 3s cooldown to eliminate voice jitter

TTS Optimization: Evaluated Chatterbox vs. commercial APIs, pivoted to Resemble AI saving 70% infrastructure costs with <200ms synthesis

GraphRAG System: Architected using Neo4j on AWS EC2, enabling multi-hop retrieval and reducing hallucinations by 25%

Deployed production on AWS Lightsail/ECS with Docker CI/CD, handling 20+ concurrent WebSocket sessions

Goonj AI: Reduced retrieval latency by 82% (1215ms→220ms) using profile-weighted retrieval with 86.7% Redis cache hit rate

Integrated AI assistant chatbot using RAG, capable of system platform-aware responses

Developed real-time opportunity finder with map integration, worldwide location detection, and auto-generated search radius circles (5-100km)

Built full-stack web solutions using Next.js, React, Node.js, PostgreSQL, and Prisma in a distributed team

Implemented RESTful APIs and optimized queries, reducing response times by 35%