Building the dojo for agentic AI.

Kojen is an end-to-end platform for training, evaluating, and continuously improving agentic AI systems. We give teams a single workflow — from problem statement to deployed model — guided by an AI trainer ("Sensei") that critiques and refines every step.

Today, fewer than 13% of ML projects reach production. Teams cobble together a dozen vendors, lose months to data plumbing, and watch deployed models silently regress. Kojen replaces this with a disciplined, opinionated dojo: dataset forge, training katas, evaluation, deployment, and continuous learning — in one system, with Sensei in the loop.

We are raising a $6M seed round to scale Sensei, expand our kata library, and bring Kojen to every team building agentic AI in the next 18 months.

Vision: A world where every product team can train, evaluate, and ship an agentic model with the same confidence they ship code today.

Mission: Build the definitive training platform for agentic AI — opinionated, guided, and continuously learning — so practitioners spend their time on judgment, not plumbing.

Our north-star metric is models in production per active team. When that number rises, every other business metric follows.

Modern AI teams operate in a fragmented stack: notebooks for experimentation, scripts for data, separate vendors for fine-tuning, evaluation, deployment, and monitoring. The result: brittle pipelines, slow cycles, and silent failure.

The cost of failure is enormous: average spend per failed model attempt exceeds $2M when you account for engineering time, GPU burn, and opportunity cost.

Kojen unifies the training workflow into a single, opinionated dojo. Practitioners describe intent; Kojen generates the dataset schema, suggests architecture, runs training, evaluates against katas, deploys, and monitors for drift — with Sensei guiding every decision.

The result: cycle time drops from weeks to hours, model quality improves measurably, and teams ship more models per quarter than they previously shipped per year.

Foundation models commoditized inference. The next wave of value is in training, fine-tuning, and continuous improvement — accessible to every team, not just frontier labs. Open weights (Llama, Mistral, Qwen), agentic workflows, and enterprise compliance demands are converging into a generational tailwind.

Sources: Grand View Research, IDC AI Infrastructure Forecast 2025, internal modeling.

Kojen ships as a unified web platform with a CLI and SDK. Every surface is built around the same primitive: a kata — a reproducible training routine with dataset, model, eval, and deploy spec.

Today's incumbents specialize in slices: data labeling (Scale, Surge), training infra (Modal, Together), MLOps (Weights & Biases), and inference (Replicate, Fireworks). None cover the full agentic loop with an opinionated trainer.

Kojen's moat: the kata library + Sensei + continuous learning loop. Each customer that trains on Kojen makes Sensei sharper, and every shared kata strengthens our marketplace.

Kojen monetizes through tiered subscriptions plus usage-based billing for compute and inference. Self-serve practitioners become team accounts, which become enterprise contracts.

We win the practitioner first. Kojen's free tier and open kata marketplace pull individual ML engineers in; viral mechanics (shared kata, public leaderboards, embedded Sensei reviews) create organic distribution. Once 2+ practitioners on a team adopt Kojen, an inbound team-tier conversation begins.

GTM headcount in year one: 1 head of growth, 2 DevRel, 1 founding AE, 1 founding SE. We expect 70% of new ARR to originate from product-led signals in the first 18 months.

Kojen runs as a cloud-native platform with multi-tenant isolation, GPU-aware scheduling, and a managed control plane. The customer surface is built on TanStack Start with edge SSR; the training plane orchestrates jobs across multiple cloud GPU providers with automatic spot/on-demand arbitrage.

Internally, we ship daily, run weekly Sensei evaluations against a held-out kata suite, and review every customer training run in a shared support channel. Operational discipline is part of the brand.

Today: 11 FTE (6 engineering, 2 AI research, 1 product, 1 design, 1 operations). Year-one plan: scale to 28 FTE, weighted toward AI research and developer relations.

Advisors and angels include operators from OpenAI, Anthropic, Hugging Face, and the AW3 venture studio network.

Our financial plan models conservative bottom-up adoption with top-down enterprise expansion in year two. We reach default-alive cash-flow at $14M ARR, with a path to $20M ARR by end of year three on a single seed + Series A.

Use of funds skews toward AI research (Sensei) and engineering; GTM ramps in late year one as product-led signal proves repeatable.

We are partnering with funds and operators who want to shape how the next decade of AI is trained. The round is led by a committed lead; we have room for a small number of additional partners with relevant networks in AI infrastructure, developer tools, or enterprise GTM.