Forge your own AI model.
KeenForge is an AI agent that distills our CTO's 7-year vision AI workflow — so anyone with trained eyes can train their own expert-grade model. No code. No cloud. You own it — and you earn from it.
What We Do
GPT-4V scores 22% on industrial defect detection. The factory needs 98%+. Generic models hit a structural ceiling in specialized domains. The people who could train expert models — QC engineers, radiologists, agronomists — are locked out by the code barrier. KeenForge replaces the entire ML pipeline with one AI agent. Expert labels and reviews. AI handles everything else — from data cleaning to deployment.
The Theory Behind KeenForge
SAI: AI Must Embrace Specialization via Superhuman Adaptable Intelligence — Goldfeder, Wyder, LeCun & Shwartz-Ziv (2025) — argues the future of AI is not one general model, but thousands of specialized superhuman systems. This is the thesis KeenForge is built on.
The Duo
.jpg)
KeenForge
Not a labeling tool. An AI vision engineer agent. Desktop app. Expert opens it → imports images → draws boxes → AI trains in the background, in real time. The full pipeline runs locally. Your data never leaves your machine.
30-50 imgs
Training
VLM
Data
LLM
Green = AI automated | White = expert involved
Hours to days per loop.
Producer-Consumer threading.
This is our early-stage prototype — the foundational HITL real-time training framework demonstrated at CVPR 2026. We are now integrating more AI capabilities into the pipeline: VLM edge-case discovery, synthetic data generation, and LLM-driven auto-tuning.
Publication
"A Curated and Re-annotated Peripheral Blood Cell Dataset Integrating Four Public Resources"
First Author — Curated the dataset, led annotation strategy, and drove the research from clinical insight to publication.