Depth

Kompetens som syns i processen

Vi säljer inte «AI-magi». Vi visar hur beslut tas: mät före merge, avstå före fel edit, behåll patentkänsligt internt.

  • Measure before merging.
  • Abstain beats a wrong edit.
  • Public pages state capability; internals stay proprietary where needed.
  • One layer per job — product vs Cursor agent vs human ops.

Eval before ship

Product changes land with gates. If the metric moves the wrong way, the default stays off.

  • ·Repo API: false_edit_rate, correct_abstain_rate, out_of_scope_handled, loop false_success / collateral.
  • ·Live e2e is reported with variance — we do not paper over unsolvables by inventing files.
  • ·Config provenance and fingerprints so “what ran” is reconstructible.

Knowing when not to act

A coding agent that always writes is a liability. Ours refuses create-repo, install, deploy, and push until those are product surfaces.

  • ·out_of_scope before retrieval for clear capability requests.
  • ·no_relevant_files when top hits sit below a measured relevance floor.
  • ·ambiguous_target / empty_edit guarantees instead of guessing.

Cognitive systems (high-level)

We build agents with memory, anomaly, and emotion signals — not only prompt chains. Implementation detail stays inside the company.

  • ·Hyperdimensional representations for robust similarity.
  • ·Active Inference as a decision frame.
  • ·Spaced retention (Ebbinghaus-style) instead of fixed reminder spam.
  • ·MCP-first integration so tools are contracts, not ad-hoc scrapes.

Azure-native delivery

Production surfaces run on Azure (SWA, Container Apps, AI Foundry) with cost awareness before new resources.

  • ·gracestack.se on Azure Static Web Apps (Standard).
  • ·BYOK / Foundry for controlled model spend.
  • ·Key Vault and managed identity preferred over scattered secrets.

Always-on lab

Max VM runs a fleet of user services — Bride, Mimer, Hermes, kodverk, STIL.AI, holidays API — so demos are processes, not slides.

  • ·systemd user units with health checks.
  • ·Public hosts where appropriate (cursor / mimer / helgdagar).
  • ·Read-only status MCP for agent introspection.

Vad vi inte publicerar

Algoritmiska detaljer för Frankenstein och viss HardwareForge-F&U (HDC+RAG) stannar internt tills IP-läge är klart. Publik yta för HardwareForge är live-produkten på hardwareforge.gracestack.se.

Se byggenInvesterare