# Jyosa Labs

> Jyosa Labs is an AI-Native Consulting Studio that designs, builds, and ships production-grade LLM infrastructure, multi-agent RAG systems, voice AI, GPU-optimized inference pipelines, fine-tuned models, and applied ML — for startups and enterprises worldwide.

Applied AI Consulting · LLM Infrastructure · Production ML Systems
Founder: Vipul Maheshwari
Email: hello@jyosalabs.ai
Website: https://jyosalabs.ai
Location: Bangalore, Karnataka, India
Book a call: https://cal.com/jyosalabs-vipulmaheshwari

## What We Are

Jyosa Labs is an engineer-led applied AI consulting firm. We build, deploy, evaluate, and operate AI systems that run in production.

## Core Capabilities

- LLM deployment and GPU optimization: vLLM, FP8/BF16/INT8, tensor parallelism, KV cache optimization, prefix caching, chunked prefill, H200/A100/L40S clusters.
- Multi-agent architecture: tool-calling agents, coordinator-routed sub-agents, DSPy pipelines, prompt optimization, and reliability guardrails.
- RAG and retrieval: semantic chunking, hybrid retrieval, LanceDB, Pinecone, Weaviate, BM25/RRF fusion, reranking, and multimodal retrieval.
- Fine-tuning and custom models: LoRA, QLoRA, FIM for code completion, domain adaptation, and custom evaluation harnesses.
- Voice AI: STT to LLM to TTS streaming, queue-based architectures, low-latency tool calling, and knowledge-grounded calls.
- Applied ML: scikit-learn, CatBoost, XGBoost, CLIP, production APIs, monitoring, and retraining.

## Primary Pages

- Home: https://jyosalabs.ai/
- Resources: https://jyosalabs.ai/resources/
- Case Studies: https://jyosalabs.ai/resources/case-studies/
- Blog: https://jyosalabs.ai/resources/blogs/
- Privacy Policy: https://jyosalabs.ai/privacy-policy/
- Terms of Service: https://jyosalabs.ai/terms-of-service/

## Proof Points

- Juspay: deployed Qwen3-VL-235B on 8xH200 GPUs with FP8, 262K context, vLLM serving, and code-model fine-tuning.
- Juspay: fine-tuned Qwen 14B FIM model reaching 52.3% exact-match accuracy and 94% token accuracy.
- LanceDB: built 8+ open-source ML/RAG applications and Lance data tooling.
- Endeavor Labs: built CatBoost labor prediction models that improved picking and packing efficiency by 15%.
- Bolna: built a knowledge-base RAG feature for real-time voice AI pipelines.

## Contact

- Email: hello@jyosalabs.ai
- Book an AI consulting call: https://cal.com/jyosalabs-vipulmaheshwari