Piyush Aaryan

Architected and built Singularity Learn’s AI Teaching Agent—a multi-model, curriculum-aware conversational teacher delivering personalized K-12 lessons with 56+ dynamically triggered interactive tools; engineered a9-mixin modular agent backend (FastAPI) with multi-provider LLM support (Gemini, Claude, GPT-4o via OpenRouter), SSE streaming, pedagogical guardrails, and AI-powered curriculum content extraction; fine-tuned Qwen Coder using RL/GRPO on a curated dataset; trained Orpheus TTS and Qwen 3 TTS on Gujarati, Hindi, and Tamil; built AI avatar streaming pipeline for real-time talking-head video with lip-sync; optimized LLM/TTS inference with TensorRT and vLLM; developed real-time voice pipeline with STT and emotion detection.

Delivered production AI backends for 6+ client applications integrating LLM workflows, diffusion inference, and scalable APIs. Implemented serverless LLM + diffusion inference pipelines (FastAPI, Celery, Redis, PostgreSQL), achieving $5,000/month in infrastructure cost savings. Shipped GPU-hosted RAG APIs using Llama.cpp + LlamaIndex + FastAPI with Cloudflare Tunnels; established end-to-end monitoring with Prometheus & Grafana. Authored multi-layer LLM safety controls covering prompt-injection resistance, policy enforcement, and safe tool usage across agent toolchains.

Delivered production AI backends for 6+ client applications integrating LLM workflows, diffusion inference, and scalable APIs (e.g., couple.me, knoetik.ai, makeitquick.ai, clarifyme.ai, Voops, Design Homes AI). Implemented serverless LLM + diffusion inference pipelines (FastAPI, Celery, Redis, PostgreSQL), achieving ~US$5,000/month in infra cost savings; shipped GPU-hosted RAG APIs using Llama.cpp + LlamaIndex + FastAPI with Cloudflare Tunnels; established observability with Prometheus + Grafana; authored multi-layer LLM safety controls.