🚀 ML Engineer @ Juspay (Xyne) | LLM Fine-Tuning & RAG Researcher | Founder – Asthra AI 💡 Building scalable, privacy-first AI systems that bridge research with real-world impact

I’m passionate about building AI systems that think, adapt, and deliver measurable value.
My work lies at the intersection of:

• 🧠 LLM fine-tuning, alignment, and evaluation 1B to 300+B
• 🧩 Efficient training (LoRA, QLoRA, DAPT, CPT, Unsloth, BitsAndBytes)
• ⚙️ RAG & Retrieval Optimization (Vespa, BEIR, custom metrics)
• 🧱 Production AI systems — from infra to evaluation pipelines

At Juspay (Xyne), I’m helping design enterprise-grade AI assistants that summarize emails, analyze docs, and reason over enterprise data.
My contributions include:

• 🧬 SFT-Play: Modular fine-tuning framework (LoRA, QLoRA-ready, 8 GB VRAM-friendly - Multi Node H200's)
• ⚡ Play / Xyne-Play: Juspay’s internal training framework for large-scale LLM evaluation
• 📊 RAG Evaluation Pipelines: Custom BEIR benchmarking, ranking, and retrieval metrics
• 🔍 Vespa Search Stack: Integration of GTE, BGE, E5, and Qwen embeddings with ranking profiles

Reusable fine-tuning environment supporting LoRA, QLoRA, Unsloth, and BitsAndBytes backends.
→ VRAM-adaptive training, backend-stamped runs, and TensorBoard logging.
⭐ 51+ stars | Used across open-source labs and internal Juspay research.

Juspay’s multi-LLM training and evaluation framework, co-maintained as open-source.
→ Integrates evaluation pipelines, multi-model routing, and deployment scripts.

A local, privacy-first AI assistant powered by RAG + multi-LLM routing.
→ 94 % document-QA accuracy; supports offline execution through Ollama.

Published IJRASET paper — hybrid ML + local LLM email assistant with self-learning loop.
→ End-to-end intelligent email summarizer, prioritizer, and replier.

• 📄 Fast-VAT: A Privacy-First Hybrid AI Email Assistant Using ML and Local LLMs – IJRASET, 2025
• 📄 MediMatch: Medicine Suggestion by Symptom Analysis – IJIRCCE, 2024
• 🧾 Fast-VAT Preprint – arXiv 2507.15904
• 🧾 "Profiling LoRA/QLoRA Fine-Tuning Efficiency on Consumer GPUs: An RTX 4060 Case Study" – arXiv 2507.15904

💻 GitHub
🔗 LinkedIn
📧 avinash.mynampati@gmail.com