About
I care about the gap between a working demo and a system that holds up in production — retrieval precision, inference latency, multilingual tokenization, and the data decisions that make models actually useful for Indian users. What I've shipped: → Production RAG chatbot for National Institute on GCP — 15,000+ queries, sub-2s latency, hybrid FAISS + BM25 retrieval with cross-encoder reranking. Optimized inference 60% (5s → 2s) via FP16 quantization and GPU batch parallelism. → Multi-agent IT/HR platform (NeuraDesk) — 4-node LangGraph system with DSPy-compiled classifier (66.7% → 93.3% accuracy), A2A protocol endpoint, and production safety gates. 111 tests, deployed on GCP Cloud Run. → Offline multilingual voice agent (Yojana Sahayak) — Whisper ASR → QLoRA fine-tuned Qwen2.5 → FAISS RAG → TTS. Zero internet dependency. MCP tool-calling server for 591+ government schemes. Model and dataset open-sourced on HuggingFace. → QLoRA fine-tuned Qwen2.5-1.5B on 39,957 bilingual Hindi-English instruction pairs — perplexity 1.15, trained in 54 min on Kaggle T4. Competitions: 🏅 Top 12 / 118 teams — NCIIPC Startup India RAG Document Retrieval Challenge 🏅 Top 3 / 100+ teams — Apollo Hospitals × IIT Delhi Medical AI Solvathon Stack: LangGraph · LangChain · DSPy · MCP · A2A · FAISS · BM25 · Whisper · QLoRA · vLLM · FastAPI · GCP · Docker Currently building GenAI and agentic AI tooling in a Managed Services capacity at PwC India — Delhi NCR. GitHub: github.com/Subh24ai HuggingFace: huggingface.co/Subh24ai
Work experience
GenAI Engineer at PwC India, Jun 2026 - Present · 3 mos, Gurugram, Haryana, India
AI Engineer at IWAYPLUS, Feb 2025 - Apr 2026 · 1 yr 3 mos, South Delhi, Delhi, India
Data Scientist at Technocolabs Softwares, Aug 2024 - Jan 2025 · 6 mos, Noida, Uttar Pradesh, India
AI/ML Engineer at CipherByte Technologies, Jul 2024 - Aug 2024 · 2 mos, Kolkata, West Bengal, India
Coding Instructor/Mentor at [E-Code], JKIAPT, May 2022 - May 2024 · 2 yrs 1 mo
