VisQuery
Visual precedent search for architects, powered by a fine-tuned vision-language model.
Style accuracy raised from 61.4% to 87.4%
Shiva Shrestha
From wireframe to working AI.
AI developer and engineer in Paris, working freelance as an AI product builder. I own products end to end: image and video generation, AI agents and skills, and design tools that take Blender and SketchUp models to photoreal AI renders.
Building visquery.com and Web Intelligence. Available for freelance AI product work and open to AI engineer roles in the EU.
Tool-calling agents, reusable skills and MCP servers that connect LLMs to real software.
What I build
- Image and video generation
- Diffusion and video-generation pipelines built for production: prompting, conditioning, fine-tuning and batch workflows.
- 3D to AI render
- Blender and SketchUp models turned into photoreal AI renders, keeping the original geometry and camera.
- Blender prompt add-on
- An MCP-style Blender extension. Describe a change in plain language and the scene updates.
- AI agents and skills
- Tool-calling agents, reusable skills and MCP servers that connect LLMs to real software.
- Design and visualisation
- Tools that help architects and designers explore, compare and present visual ideas faster.
- Idea to product
- Scoping, prototyping, evaluation and deployment, from a first sketch to a shipped product.
Work
Web Intelligence
Production RAG website analyst with a custom evaluation harness.
Eight quality metrics, reported openly
Blender Prompt Bridge
Blender add-on that connects an LLM to the scene through an MCP-style tool protocol. Type a prompt and objects, materials, lights and cameras change.
Blender Python API, tool calling
Private code, demo on request
3D to AI Render
Takes Blender and SketchUp models and produces photoreal AI renders, conditioned on the original geometry and camera.
Blender, SketchUp, controlled diffusion
Private code, demo on request
Generative Image and Video
Generation workflows for design and marketing: prompt to image, image to video, and one consistent style across a batch.
Text to image, image to video
Private code, demo on request
AI Agents and Skills
Tool-using agents and a library of reusable skills that turn a plain-language brief into working software and multi-step workflows.
Agents, skills, MCP servers
Private code, demo on request
DAM Multimedia Gallery
Digital asset manager with AI search over images and media, backed by a FAISS vector index.
Next.js, FastAPI, SQLite, FAISS
Writing
- The 2026 European Heat DomeWhy our concrete cities are failing, how architecture must fight back, and how AI can predict and optimize the best ways to stay cool.VisQuery JournalJul 2026
- Fine-tuning a VLM is mostly not a training problem. Here are the five decisions that mattered more.MediumMay 2026
- What I Learned Measuring My Own RAG SystemMediumMay 2026
- Fine-tuning CLIP on a Niche Domain: How I Got +26pp Accuracy on Architectural Stylesdev.toMay 2026
- Building a RAG Evaluation Harness That Actually Catches Problemsdev.toMay 2026
Background
Today I work freelance as an AI product builder, owning each product from idea to launch: generative image and video systems, AI agents and skills, and 3D-to-AI-render tooling for Blender and SketchUp, including a prompt-driven, MCP-style Blender add-on. Previously at AREP in Paris, where I built CLIP-based multimodal search and image-tagging systems used by internal teams. Before that, three years building ETL pipelines and large-scale web scraping at Procit, and earlier work annotating radar and LiDAR data for autonomous driving perception models at CloudFactory. MSc in Computer Science (Data Science & Analytics) from EPITA Paris, 2025.