Divax Shah

Building |

AI/ML Engineer crafting premium-grade solutions in LLM fine-tuning, NLP, Computer Vision, and Generative AI. Passionate about bringing ideas to life with precision and elegance.

View ProjectsDownload Resume
FIND ME ON
🤗
Scroll

Pushing the boundaries of what AI can do.

I'm an AI/ML Engineer with a deep passion for building intelligent systems that solve real-world problems. From fine-tuning 7B parameter models on Sanskrit to building multilingual AI detectors that rival industry leaders, I thrive at the intersection of research and engineering.

Currently at Avinyaa Edtech, building next-generation AI writing tools. Previously at Thinkbiz Technology and DMI Finance, where I shipped production-grade LLM pipelines and synthetic data systems.

3+
Years in AI/ML
12+
Projects Built
95%
Best Model Accuracy
3
Companies

LLM Fine-tuning

Specialized in adapting large language models with LoRA, QLoRA, and Axolotl for domain-specific tasks.

Multilingual AI

Built multilingual systems covering English, Hindi, Spanish, German, Dutch, Italian, and Sanskrit.

Production Systems

Delivered scalable AI pipelines from PoC to production, improving accuracy by up to 6× over baselines.

Community & Open Source

Active contributor to the AI community with 10+ public models on HuggingFace and tools for Sanskrit NLP.

Where I've worked

Jr. AI/ML Developer

Current
Avinyaa Edtech Private Limited
Mar 2025 – Present
India
  • Developing and refining an advanced grammar checker for KreativeSpace's AI Writing Tools Suite by fine-tuning LLMs for high accuracy and natural language output.
  • Designing and training a multilingual AI text classification system to detect AI-generated content in English, Hindi, Spanish, German, Dutch, and Italian — achieving up to 95% accuracy, often outperforming Quillbot's detector and approaching GPTZero.
LLM Fine-tuningNLPMultilingual AIGrammar AI

Jr. Python Developer

Thinkbiz Technology Private Limited
May 2024 – Mar 2025
India
  • Built a scalable PoC of an advanced chatbot system with LangChain Agents and LLM, handling structured (SQL) and unstructured data with ~17 agents, intercommunication, RBAC, real-time ingestion, and automated source citation.
  • Developed an OCR–LLM pipeline for Jugaad, improving invoice extraction accuracy from 15% → 85% and receipts from 10% → 90% across formats.
LangChainLLMOCRRAGPython

AI & Synthetic Data Developer Intern

DMI Finance Private Limited
Jan 2024 – Apr 2024
India
  • Developed a generative AI system for synthetic structured data generation with Python + Gradio for cleaning, deduplication, and embedding.
  • Built a robust PyTorch + Transformers fine-tuning framework enabling LLM adaptation across diverse datasets, and delivered a user-friendly Gradio interface for synthesis.
Generative AIPyTorchGradioSynthetic Data

Things I've built

Model

Sanskrit Translate v2

Fine-tuned Qwen2.5-7B-Instruct optimized for Sanskrit language processing. Features Sanskrit to IAST transliteration, bidirectional Sanskrit ↔ English translation with context-aware preservation. Enhanced with specialized dataset, chat template format, and optimized LoRA configuration.

LLMFine-tuningSanskritQwen-2.5-7BLoRA
Model

Sanskrit Tokenizer

Native Sanskrit-English tokenizer for Qwen2.5 providing 4.5× better efficiency than byte-level tokens. Produces clean, readable tokens with 120K vocabulary trained on massive English+Sanskrit corpus.

TokenizerNLPSanskritBPEHuggingFace
Model

Sanskrit Translate v1

Fine-tuned Qwen2.5-7B-Instruct-1M on Sanskrit datasets using QLoRA (Axolotl), gradient checkpointing, 4-bit quantization, and cosine LR to translate Vedic Sanskrit → English.

LLMQLoRASanskritAxolotl4-bit
Website

Flux LoRAs

Flux.dev model fine-tuned on different images creating various LoRAs that you can explore. Specialized LoRA adapters for different visual styles and concepts.

Generative AIImage GenLoRAFluxFine-tuning
Model

Geolocation via Image Classification

Transfer learning with VGG16 to identify Indian cities from images, achieving 66.3% accuracy. A computer vision project exploring geographic image understanding.

Computer VisionDeep LearningTransfer LearningVGG16
Model

Character Chatbot

Conversational agents with DialoGPT-medium: interactive chats as Tony Stark, Harry Potter, and more. Fine-tuned for character-consistent dialogue generation.

LLMChatbotFine-tuningDialoGPT

My toolkit

Core Languages
PythonTypeScriptSQL
ML Frameworks
PyTorchTensorFlow / Kerasscikit-learnHuggingFace TransformersAxolotlUnsloth
AI Domains
LLM Fine-tuningGenerative AINLPComputer VisionPrompt EngineeringReinforcement LearningSynthetic Data
Orchestration & Tools
LangChainLlamaIndexGradioFastAPIAWS
APIs & Services
OpenAI APIGoogle Gemini APIAnthropic APIMistral AI APIGroqOpenRouter
Data & Libraries
NumPyPandasOCR PipelinesRAG SystemsVector Databases
Python·TypeScript·SQL·PyTorch·TensorFlow / Keras·scikit-learn·HuggingFace Transformers·Axolotl·Unsloth·LLM Fine-tuning·Generative AI·NLP·Computer Vision·Prompt Engineering·Reinforcement Learning·Synthetic Data·LangChain·LlamaIndex·Gradio·FastAPI·AWS·OpenAI API·Google Gemini API·Anthropic API·Mistral AI API·Groq·OpenRouter·NumPy·Pandas·OCR Pipelines·RAG Systems·Vector Databases·Python·TypeScript·SQL·PyTorch·TensorFlow / Keras·scikit-learn·HuggingFace Transformers·Axolotl·Unsloth·LLM Fine-tuning·Generative AI·NLP·Computer Vision·Prompt Engineering·Reinforcement Learning·Synthetic Data·LangChain·LlamaIndex·Gradio·FastAPI·AWS·OpenAI API·Google Gemini API·Anthropic API·Mistral AI API·Groq·OpenRouter·NumPy·Pandas·OCR Pipelines·RAG Systems·Vector Databases·

Let's build something remarkable.

I'm always open to discussing AI projects, collaborations, or just having a great conversation about the future of technology. Reach out through any of the channels below.

GitHub
@shahdivax
See my code & projects
LinkedIn
divax-shah
Connect professionally
X (Twitter)
@divax_shah_
Follow my AI journey
🤗
HuggingFace
diabolic6045
Explore my AI models
Email
divax12345@gmail.com
Drop me a message
2026 Divax Shah.
Built with Gemini AI
Available for opportunities