You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
ServiceWhy: an agent that explains why a restaurant night went wrong. Planner, Investigator, Verifier and Writer over Zetaris federated SQL, Meterless World Model and Nightshift. No receipt, no claim. Synthetic data.
Presensi School - Smart attendance system with Face ID, GPS, and AI agent. Data layer: Zetaris. AI: Meterless Nightshift (Nemotron). Full stack: PHP + Python + Docker.
Ask why a number moved: four agents investigate revenue changes across an ERP and CRM federated in Zetaris, cite the SQL behind every claim. Nemotron via Meterless Nightshift, memory in Meterless H-MEM. Open Agent Hackathon 2026, Track 2.
AI growth agent for indie app developers: proposes prioritized growth actions with SQL-backed claims, checked by an independent Critic. Built with Strands Agents SDK, Zetaris, Meterless Nightshift / H-MEM, and NVIDIA Nemotron.
An AI agent that explains why a small shop's profit dropped. It checks sales, bank, supplier and expense records through Zetaris, and shows the records behind every cause.
Platelets that reach the right hospital on time. Reads every hospital's live stock through Zetaris and never moves a unit its own record no longer supports.
Agents that explain why gross margin moved, with every claim traced to rows in GL, AP, sales and bank through Zetaris. Open Agent Hackathon 2026, Track 02.
AI investigator for product-safety defects: finds the root cause, proves every other explanation wrong with stratified statistics, and quarantines only the affected units. Live at traceon.netlayerlabs.com. NVIDIA Nemotron via Meterless Nightshift, Zetaris, Meterless H-MEM.
Ask why your electricity bill changed. An agent that splits the change exactly, tests 11 causes on Zetaris-federated data, and answers with NVIDIA Nemotron via Meterless Nightshift; every figure traced to its query. Open Agent Hackathon 2026, Track 02.
Research that works like a consulting team: break a hard question into an issue tree, staff every branch, and hold every number to a stored, checked source. Code, not a model, decides release.