hiloop

Hiloop operates an autoresearch service in which a customer supplies a measurable target and a scoring method, and the system converts that objective into thousands of directed experiments across data, post-training, continual learning, prompts, tools, and systems. The engagement concludes with a verified improvement delivered as a better model or algorithm.

Engagement formats include post-training against customer-owned evaluations, inference and kernel optimization, synthetic reinforcement learning environments, and signal processing over labelled corpora. Campaigns run on hosted infrastructure or inside the customer cloud, and persistent memory, experiment lineage, and statistical verification are built into the service.

The service addresses any team able to write down a metric and afford to evaluate it: small post-trained models, latency-sensitive applications, and tunable systems where a score exists. The company frames that boundary as the limit of what its campaigns can reach, since an agent can only climb a target that is measurable.

As individual experiments grow cheaper, the accumulated record of campaign results compounds into a memory base and training corpus for later research agents. That compounding effect gives the loop an advantage that grows with each engagement rather than resetting, which is the structural bet behind the company's market position.

Hiloop publishes benchmark evidence for its campaign approach: a validation bits-per-byte of 0.9016 recorded as the median of 25 confirmation runs on an autoresearch benchmark published by Andrej Karpathy, against a published state of the art of 0.9109 on B200 hardware. The campaign reached that figure with unmodified stock coding agents, so the gain is attributed to the orchestration layer rather than custom models.

The same system implemented NanoGPT with a runtime 12.54 percent below the upstream reference and executed 4,188 experiments across two days on 50 B200 accelerators. Together those results indicate that breadth and depth of automated search scale without hand-tuned pipelines, which is the differentiator the company builds its service on.

Hiloop runs with a two-person team and lists no open roles, so campaign capacity currently depends on the founders own bandwidth. The founders describe autoresearch as a field still in its infancy, and the competitive field includes named laboratories pursuing recursively self-improving systems built on the same loop.

The product surface thins at the edges: default setups track experiment state in a markdown file and a results table, an approach that strains at a thousand parallel experiments, and the company publishes no pricing page, routing prospective customers through a founder email address. Those constraints shape how quickly the service can scale beyond early engagements.

Hiloop lists no pricing page and no tiers, and the pricing route returns no page, so commercial terms are reached through direct conversation. The homepage directs visitors to talk to the founders or to join an email list for updates rather than to a checkout flow.

Campaigns can run on hosted infrastructure or inside the customer cloud, which keeps deployment flexible while the commercial motion stays founder-led. This is a direct-sales posture typical of an early-stage technical services company still calibrating what to charge.