Vrushank Dabhi

AI/ML Engineer with 1.5+ years of experience building production-ready Generative AI, LLM, RAG, and Document AI solutions. Experienced in developing end-to-end AI systems using Python, PyTorch, LangChain, LangGraph, FastAPI, Qdrant, and vLLM, with hands-on experience in LLM inference, model serving, prompt engineering, evaluation, and AI agents.

I have worked on scalable AI applications involving LLM/VLM integration, intelligent document processing, OCR, vector search, structured information extraction, and agentic workflows. I enjoy solving complex problems and turning AI/ML research concepts into reliable, production-oriented systems.

Currently seeking opportunities as an AI Engineer, ML Engineer, GenAI Engineer, or AI/ML Backend Engineer, where I can contribute to building impactful AI products and continue growing in the field.

AI/ML Engineer at Trestle Labs (2024-10 – Present)

• Kibo Studio: Architected the FastAPI backend for Kibo Studio's document-intelligence and data-extraction modules, OCR digitization, table-to-Excel and multi-format document transforms, and multimodal PDF/image-to-JSON extraction across ten regulated domains (banking, insurance, healthcare, KYC, legal, etc) using Google Cloud Vision, Gemini, and Claude. Built a shared platform layer with Bearer API-key auth, per-feature entitlements, atomic credit reserve/refund over DynamoDB, and S3-backed artifact storage. Engineered async orchestration with bounded thread-pool concurrency for inference-heavy pipelines, YAML-templated prompt versioning with page-level processing, and structured observability.
• Newspaper Ads Collection-An AI Powered Solution: Designed a multi-stage pipeline for similar ads image classification, involving: image segregation using CNN model between display and no display images, generating image embeddings using CLIP and text embeddings using Labse model, Doing OCR for non display images and key feature extraction for display images, indexing and storing embeddings in Qdrant, calculating the cosine similarity score between combination of embeddings+OCR/key features, classifying similar images using a Random Forest regression model and Knee-Elbow method thresholding to yield relevant similar image sets. Generating excel file for each each with similarity score indicating if images are similar or not.
• Production LLM serving backend: Built and shipped a type-hinted, async FastAPI service that serves structured JSON extraction over REST from multi-page PDFs. Backed by Qwen3.5 on vLLM or Gemini API call with a two-stage inference flow; implemented semaphore-limited async concurrency, single + batch (ZIP) ingestion, and ETL/normalization over unstructured documents.

BTech in Electrical Engineering – Nirma University (2021 – 2025)