DEW · a foundation model for education

Rainwater Labs is an artificial intelligence research and foundation model company.

AI engineering is currently outpacing AI science. We are scaling frontier models faster than we understand their underlying mechanics, leaving deep vulnerabilities unresolved.

Because core training remains centralized, organizations are locked into dangerous platform dependencies rather than owning their intelligence. Worse, these massive systems are structurally rigid. Despite their raw power, they are notoriously difficult to air-gap, deeply customize, or force into alignment with strict, specialized compliance protocols.

Rainwater Labs exists to engineer foundational architectures designed to be fundamentally understood, precisely customizable, and highly capable in secure environments. We are a lean team of scientists, engineers, and builders who have spent decades pioneering deep-tech platforms, scaling complex infrastructure, and filing core architectural patents. We like to take on the hardest technical challenges because solving them at the foundation unlocks progress for the many.

Most AI hands students the answer. DEW doesn't. Grounded in the Socratic method, it guides children to discover answers through their own reasoning — asking structured, sequential questions until a concept genuinely clicks. It's free, open-source, and built as a public good rather than a commercial product.

API

How it works

Speaks a child's language

Pre-trained on billions of educational tokens — K-12 textbooks, curriculum guides, and public-domain literature — so its vocabulary and pacing fit the student, not a corporate user.

Teaches, doesn't answer

The Socratic method is baked into the model's weights. For students under 16, DEW refuses to dump finished answers — it asks guiding questions instead.

Safe at the weight level

A runtime guardrail clamps unsafe topics out of existence — the model is physically incapable of expressing them. No fragile prompt filters to jailbreak.

By the numbers
3B
Parameters — a light footprint tuned end-to-end, not a prompt wrapper on a corporate API.
<16
Age below which DEW hard-refuses to dump finished answers, guiding Socratically instead.
−∞
Logit clamp on forbidden concept classes — the network is physically incapable of expressing them.
Apache 2.0
Weights, code & datasets — white-label, self-host, and modify with zero restrictions.

Training pipeline

01 · Continued Pre-Training (CPT)

Initializes from a clean 3B baseline, then deep-trains on billions of educational tokens — K-12 textbooks, peer-reviewed curriculum guides, and public-domain literature — shifting the core language structure to age-appropriate vocabulary.

02 · Direct Preference Optimization (DPO)

Locks in the Socratic process over thousands of "chosen vs. rejected" pairs — chosen: multi-turn guiding inquiries, contextual hints, conceptual milestones; rejected: direct answers, final sums, passive essay generation.

03 · Logits-Level Security

A custom runtime LogitsProcessor identifies forbidden conceptual classes and clamps their generation probability to −∞ — bypassing brittle system-text blocks entirely.

Deployment

Runtime

Runs at ultra-low latency on standard IT-lab desktops, Chromebook servers, tablets, or offline classroom networks via vLLM / Ollama — no hardware upgrades required.

Data isolation

Local-first pipeline. Inputs are sanitized of PII (names, phone numbers) before inference — zero student metadata leaks to third-party corporate entities.

Research

DEW: Reclaiming Classroom Thinking in the AI Era

By the DEW Team · June 2026

Our mission brief and technical architecture — how the Socratic method is built into the model's weights, the logits-level guardrails that keep it safe, and why we ground all of it in Responsible AI.

Read the paper ↗
Our vision

It will be free forever

A child's mind should never be a business model. We intend to keep DEW free, forever — never gated behind a paywall, never sold. Our code, datasets, and methods stay open, always, under Apache 2.0, so anyone can inspect, host, and improve it.

This is how we deliver on our purpose of Responsible AI: not as a compliance exercise, but as the engineering philosophy that guides every model we build.

FAQ
Is DEW really free?+
Yes — free, forever. DEW is a public good. We will never gate it behind a paywall, charge schools, or sell access to a child's learning.
What ages is DEW built for?+
DEW is tuned for K-12 learners. For students under 16 it refuses to hand over finished answers, guiding them with Socratic questions instead; vocabulary and pacing adapt to grade level automatically.
Can we run DEW on our own infrastructure?+
Yes. DEW is a light 3B model you can host entirely in-house via vLLM or Ollama — on standard lab desktops, tablets, or even fully offline. Student data never has to leave your network.
Is it open-source? What's the licence?+
Fully open under Apache 2.0 — code, datasets, and methods. You can inspect exactly how it works, white-label it, and customise the codebase.
How does DEW keep student data private?+
Inputs are sanitized of names, numbers, and other PII by design. There are no ads and no data harvesting, and zero student metadata leaks to third parties.
Get involved

We are open to partner

DEW is free and open-source. We grow it through partners who share the mission.

Bring DEW to your school

Roll it out for your students and teachers — we'll help you set it up.

Support through CSR funds

Fund free, responsible AI education at scale through your CSR programme.

Help with research

Collaborate on the model, studies, and safety work behind DEW.