NoaAI
Multi-agent conversational and voice AI platform: centralized orchestration with strict phase isolation, RAG-grounded answers, persistent state, and real-time voice with interruption handling.
AI in simple terms
Khushal Jethava
Machine Learning Engineer
I build production Generative AI, multi-agent LLM systems and real-time voice AI from Ahmedabad, India, and I tell the story the way I learned it, through the wisdom of Shri Krishna.
- ॐThe first sound. Every story begins here.
- पाञ्चजन्यKrishna's conch, blown to open the battle at Kurukshetra (Gita 1.15).
- गाण्डीवArjuna draws the Gandiva, the divine bow given to him by Agni.
- मुरलीधरKrishna of Vrindavan, the flute-player crowned with a peacock feather.
- गीता उपदेशAs Arjuna's charioteer, beneath the Hanuman flag, Krishna speaks the Gita (Gita 1.20).
Chapter I · बांसुरी
Every great model begins by listening.
Krishna's flute is hollow. It makes music only when breath flows through it. A model is the same: nothing at all until data flows through it, shaped with care.
I'm a Machine Learning Engineer with 4+ years of designing and deploying production-grade Generative AI: LLM-powered applications, multi-agent orchestration, RAG pipelines, real-time voice AI, computer vision and low-latency AI infrastructure. I own the full lifecycle, from research and modeling to architecture, deployment, monitoring and optimization.
Along the way: Document AI with fine-tuned transformers, room transformation with ControlNet, car damage detection with Detectron2, generative chatbots with LSTM and BART, voice agents answering real phone calls, and now computer vision, AI automation and multi-agent workflows at Bytes Technolab.
4+Years in AI/ML
1000+Daily AI voice calls handled
99.5%Voice platform uptime
180+Technical articles
Chapter II · मोरपंख
A thousand eyes, one vision.
Every eye on the peacock feather sees something different, yet it is one feather. My toolkit spans language, vision and generation, held together by one goal: AI that actually works in production.
- Python
- Generative AI
- LLMs & fine-tuning
- RAG systems
- LangChain
- AI agents
- Multi-agent systems
- LangGraph
- LangFuse
- Voice AI (STT · LLM · TTS)
- SIP, Twilio & PSTN
- Embeddings & prompt engineering
- NLP
- Transformers
- PyTorch
- TensorFlow
- HuggingFace
- QLoRA & PEFT
- XGBoost & classic ML
- Computer vision
- YOLO & OpenCV
- Detectron2
- Stable Diffusion & ControlNet
- FastAPI
- Django
- Flask
- WebSockets & SSE
- PostgreSQL & Redis
- FAISS, Pinecone, Qdrant, LanceDB
- Docker
- AWS (S3, EC2, Lambda, SageMaker)
- MLOps
Chapter III · सुदर्शन चक्र
Precision, in motion.
The Sudarshan Chakra never stops turning and never misses its mark. Each role I've held sharpened the aim.
Senior Machine Learning Engineer · Bytes Technolab
Computer vision, Generative AI, AI automation and multi-agent workflows.
Machine Learning Engineer · Codiste Private Limited
- Architected a production multi-agent orchestration platform on LangGraph, with deterministic execution and strict phase isolation.
- Shared persistent state and unified confirmation logic cut redundant user interactions by 35% and lowered API costs.
- Real-time voice pipelines (STT, LLM reasoning, TTS) with sub-800ms end-to-end latency.
- OpenAI, ElevenLabs and telephony (SIP, Twilio, PSTN) handling 1K+ daily calls at 99.5% uptime.
- Advanced GenAI for deepfake technology and image-to-avatar transformation, plus production image segmentation.
Machine Learning Engineer · Maknon Private Limited
Image segmentation of floors, walls and room components; QLoRA fine-tuning to cut GPU memory; ControlNet room transformations and backend APIs serving the models.
AI Software Developer · SoluSoft Technologies
Led Document AI, computer vision and LLM chatbot work: a Document AI platform for enterprise clients (Django, FastAPI); barcode detection with YOLO and OpenCV; car damage detection with Detectron2; generative chatbots with LSTM and BART.
Winner, Codiste Hackathon 2025
Chapter IV · विश्वरूप
The whole universe, inside one form.
On the battlefield, Arjuna saw everything that exists inside a single form. A neural network is a small echo of that idea: millions of parameters, one coherent mind. These are the minds I've built.
Dialora.ai
Real-time AI calling agent: low-latency STT, LLM and TTS over SIP, Twilio and PSTN, with fallback logic, cost-aware routing and conversation recovery.
dialora.aiNdeck.io
AI proposal and video generation: an LLM engine that turns structured inputs into full technical proposals, with auto-generated diagrams, voice narration and lip-synced video.
ndeck.ioDocument AI platform
Classification and information extraction on fine-tuned transformers, served with Django and FastAPI.
Room transformation
Interior segmentation plus high-quality generative redesigns driven by ControlNet.
Car damage detection
Automated vehicle damage assessment with Detectron2 and CNNs.
Barcode detection
High-throughput barcode detection in documents with YOLO and OpenCV.
Generative chatbots
Conversational AI with LSTM sentiment analysis and BART dialogue summarization.
Image-to-avatar GenAI
Generative pipelines for deepfake technology and image-to-avatar transformation.
Chapter V · कर्मयोग
The whole cosmos, resting in His palm.
कर्मण्येवाधिकारस्ते मा फलेषु कदाचन ।
“You have a right to your work, never to its fruits.” Bhagavad Gita 2.47
No planet chases applause. Each one keeps its own orbit and does its own work, and together they hold the universe in balance. That is karma yoga.
It's also how I treat the craft: do the work well, then give it away. 180+ free tutorials on AI, LLMs and Python, each with hands-on code you can run today.
Fix 'LLM API Error: An Error Occurred During Streaming'
What causes the 'LLM API error: an error occurred during streaming' message and how to fix it — timeouts, token limit...
Best Open Source LLMs for Coding and General Use in 2026
How to choose the best open source LLM in 2026 — the top model families for coding and general use, how to match size...
LLM API Pricing Compared: How to Find the Cheapest Provider
How LLM API pricing works and how to find the cheapest provider — the per-token model, hidden costs, and a Python cos...
Epilogue · कमल
Let's build something that blooms.
The lotus rises clean out of muddy water. Bring me the messy data and the hard problem. I'm open to Generative AI, multi-agent, voice AI and computer vision work.