Teaching an AI to solve brand-new puzzles it has never seen before by writing its own mini-programs on the fly, instead of memorizing millions of examples.

Ndea's core engineering effort centers on Latent Program Networks (LPNs), a novel architecture that uses deep learning to guide the search through a space of discrete, interpretable programs. Unlike traditional large language models that rely on massive datasets and pattern matching, LPNs learn to synthesize small, composable programs that explain observed data from just a handful of examples. This enables few-shot generalization—the ability to adapt to entirely new problem types without retraining. The system leverages neural networks to propose candidate program structures, then uses symbolic execution and verification to validate correctness, creating a tight feedback loop between learned intuition and logical reasoning. This approach directly targets the ARC-AGI benchmark, which measures fluid intelligence and abstraction rather than memorization, and is designed to generalize to real-world scientific discovery tasks such as hypothesis generation, experimental design, and materials science optimization.

It's like hiring a scientist who doesn't just memorize every textbook but actually understands the underlying principles well enough to invent new theories from a single surprising experiment.

Building an AI brain for robots that can figure out how to handle objects and navigate spaces it has never encountered before by imagining what will happen before it acts.

Ndea's applied AI research arm is developing neuro-symbolic world models that combine learned perceptual representations with structured, symbolic reasoning about physical environments. Job postings for applied AI researchers with expertise in robotics, computer vision, spatial AI, and embedded systems reveal a concerted push toward embodied intelligence. These world models allow an agent to simulate the consequences of actions in a latent space before executing them physically, dramatically reducing the need for costly real-world trial-and-error. By integrating program synthesis with reinforcement learning, the system can generate novel action plans for tasks it has never been explicitly trained on—such as manipulating unfamiliar objects or navigating new environments. This positions Ndea to extend its AGI research beyond software benchmarks into real-world physical domains, potentially enabling breakthroughs in autonomous manufacturing, laboratory automation, and robotic scientific experimentation.

It's like giving a robot the ability to play out a chess game in its head before moving a single piece—except the chess board is your entire kitchen and the pieces are random objects it's never seen.

Creating an automated testing system that measures whether an AI can actually think and invent—not just parrot back answers—so the entire field knows who's really making progress toward AGI.

Through the ARC Prize Foundation and Ndea's internal research, the team is building an automated, scalable platform for evaluating general intelligence in AI systems. This goes far beyond traditional ML benchmarks that measure accuracy on static datasets. The platform uses procedurally generated, novel tasks that test fluid reasoning, abstraction, and creative problem-solving—capabilities that current LLMs struggle with. Machine learning is used both to generate diverse evaluation tasks (ensuring no system can game the benchmark through memorization) and to analyze submission patterns, detecting overfitting, data contamination, and shortcut learning across thousands of competing models. The platform also incorporates meta-learning analytics that track how the global research community's approaches evolve over time, providing Ndea with unique strategic intelligence about which hybrid architectures are most promising. This dual role—public benchmark steward and private research accelerator—gives Ndea an unparalleled feedback loop: they set the standard for AGI measurement while simultaneously using the resulting data to guide their own research priorities.

It's like being both the professor who writes the final exam and the student who gets to see how every other student in the world answers it before writing their own thesis.