At Top Notch Infotech | Ziqra AI, I built a speech-analysis pipeline for Ziqra AI, an AI Interview Coach. It combines Whisper transcription, wav2vec2 phoneme recognition, and speech features so the system can assess pronunciation, fluency, pauses, intonation, and engagement.
I fine-tuned two specialist models with LoRA and QLoRA, then deployed them with vLLM on Modal; score-only generation cut model latency by up to 7.5x. At Codincity, I built an AP/AR knowledge-graph MVP with Neo4j and LangGraph, and developed a separate anomaly-detection pipeline for accounts receivable.

