Hákon Freyr

Hákon Freyr Gunnarsson

AI engineer. I build AI systems that can be trusted with real work.

StatusOpen to full-time roles

Drawing index
SheetTitleKey dimension
02Krates4,913 checks, 0 differences
03The Weave106 merged changes / week
04Homegrown Hero Films60 fps
05ConceptStudio10,004 steps for US$3.99
06Experience6 organisations
Key plan: isometric drawing of the Krates appliance, with locators for each sheet0203040506

Cover sheet · Sheet 1 of 6 · Rev 2026-10 · Drawn H.F.G.

Krates

A seven-month build sprint, March to October 2026, that took agent-first back-office software from design to an independently verified working core. It is correct by construction: agents do the work, rules are executable logic, and a human consents before anything leaves the box.

14 services in five credential-isolated containers

7 agent-first kernels

Exploded isometric drawing of the Krates appliance: lid, cooler, memory, board, drive and chassis pulled apart along assembly lines
Sókrates
The on-premises AI department: one NixOS appliance that ingests a company’s systems and lets named agents do back-office work over Slack, Teams, WhatsApp and email.
Krepis
Seven back-office kernels (accounting, payroll, workforce management and more) built for AI agents rather than for people clicking through screens.
Revisions
RevDescription
2026-10Build phase complete. Krates continues as a side company, with development carried mainly by the agent team.

General notes

  1. Run by agents, unsupervised: AI agents operated a fictional 12-employee company’s whole month through all six Krepis kernels with no human supervisor. Payroll, 14 payments and the books came out exact: 4,913 independent checks, zero differences. Along the way the kernels rejected 11 invalid calls from the agents and wrote nothing for any of them.
  2. Business rules are Logica laws compiled to DuckDB SQL; each returns the set of violations, not a pass or fail.
  3. An agent that dies mid-task never repeats an external action (DBOS, write-ahead attempt records).
  4. Every governed action carries a signed consent token and is its own trace span in a queryable audit store.
  5. Krepis kernels are event-sourced and append-only; every action is idempotent and has a dry run.
  6. Two editions: a gateway-routed mini-PC, or fully local inference on a DGX Spark for air-gapped sites.

Detail: SókratesDetail: Krepis

Project
Hákon Freyr Gunnarsson
Sheet
2 / 6
Title
Krates
Scale
N.T.S.
Rev
2026-10

The Weave

The engineering organisation that builds Krates: five named AI engineers (Claude and Codex), each with its own machine and role, working under a written doctrine. I designed it and lead it, in the role a CTO plays.

106 merged changes a week, July to mid-September 2026

94% independently reviewed before merge (September)

  1. 01Ticket
  2. 02Write
  3. 03Review
  4. 04Repair
  5. 05Merge

250 ISK ($2) of AI subscription per merged change

Reviews always come from an agent other than the author. Repairs go back through review before the merge gate.

General notes

  1. Reviewers found issues in 83% of the changes they reviewed, and only 0.14% of merged changes have ever been reverted.
  2. A scheduler I built gives each task to the agent best placed to take it, by remaining subscription capacity, skills, availability and parallelism.
  3. Mistakes become procedure: a daily harvest turns what reviewers catch into 40 standard operating procedures holding 414 recorded lessons.
  4. A self-built broker with outbound-only edges over Tailscale wakes the right agent on the right machine from Slack, Linear and GitHub.

Detail: The Weave

Hive

The coordination fabric behind the Weave. Hive Attention brings a decision into Slack and returns the answer to the agent’s work context.

Detail: Hive

Project
Hákon Freyr Gunnarsson
Sheet
3 / 6
Title
The Weave
Scale
N.T.S.
Rev
2026-10

Homegrown Hero Films

Short films, photoreal or animated, where a child is the hero of their own adventure, made with an AI video pipeline I built end to end.

Still from Lia at Lanternfall
Lia at Lanternfall
Still from The Hangar Call
The Hangar Call
Still from Gromble and the Failed Ornithopter
Gromble and the Failed Ornithopter

60 fps finishing: NVIDIA RTX VSR upscaling and RIFE interpolation, shot by shot

General notes

  1. Generation runs on rented GPUs with my own open-source extensions to the MiniMax H3 video model.
  2. Every film is reviewed frame by frame by an AI quality gate before a human sees it.
  3. Interpolation never blends two shots: cuts are detected and every shot is finished on its own.

homegrownherofilms.comModel extensions on GitHubDetail: Homegrown Hero Films

Project
Hákon Freyr Gunnarsson
Sheet
4 / 6
Title
Homegrown Hero Films
Scale
N.T.S.
Rev
2026-10

ConceptStudio

A native Mac app that teaches an image or video model a new concept from your own photos or clips: it writes the training captions with a language model of your choice, rents a GPU within the budget you set, trains, and brings the adapter back. Development preview, October 2026.

9 screens, from a folder of photos to a trained adapter

26 trainable models in one shared table

  1. 01Images
  2. 02Concepts
  3. 03Mapping
  4. 04Captioning
  5. 05Review
  6. 06Publish
  7. 07GPU pods
  8. 08Training
  9. 09Monitor

US$3.99 of rented GPU for a 10,004-step run through the app (14 October 2026)

The Mac does the preparation; a GPU computer the app rents and controls does the training, inside the budget and idle limit set before renting.

General notes

  1. One model table, two readers: the app builds its pickers and limits from it, and my GPU images write the trainer’s configuration from the same file.
  2. Captions are shaped for the model being trained by a swappable Agent Skill, written by any language model chosen at run time.
  3. Rented computers run my own container images; the app starts services from a fixed catalogue over one authenticated API, never from shell strings.
  4. A likeness adapter on Qwen-Image-2.1 went from 0.05 to 0.34 face match to held-out photos in 2,500 steps (15 October 2026).

Code on GitHubDetail: ConceptStudio

Project
Hákon Freyr Gunnarsson
Sheet
5 / 6
Title
ConceptStudio
Scale
N.T.S.
Rev
2026-10

Experience

Mathematics and statistics first, then production systems in genomics, finance, travel and enterprise software, each a new domain learned within months.

Parts list
ItemOrganisationRolePeriodKey result
01Krates ehf.Founder & Lead EngineerMar 2026 – PresentBuild sprint to a verified core: 4,913 checks, zero differences
02Wise lausnir (Microsoft Dynamics partner)Backend EngineerOct 2024 – Mar 2026Fabric onboarding cut from about three weeks to half a workday
03TravelshiftData Scientist & Business Intelligence LeadAug 2022 – Sep 20240.42% forecast error on 2.4B ISK of cash flow, six months out
04Alfreð AtvinnuleitData Analyst & Product OwnerMay 2021 – Aug 2022Grant proposal that secured 30M ISK from Rannís
05Reykjavík UniversityInstructor, Department of Computer Science2018 – 2021Taught programming, data structures, calculus and statistics
06deCODE Genetics (Amgen)Statistician & Bioinformatician2015 – 2017Haplotype compression more than 10× better than gzip
Related drawings
HiveCoordination for Agents. Attention for a Human
Residual ModelingInterpretable, Counterfactual Valuation of Icelandic Homes
Memory PalaceGraph-Based Semantic Memory System
Grimoire (SETS)Enterprise Knowledge Management Platform
AutopodMulti-Tenant Podcast Orchestrator
Sókrates IDRLegal Document Intelligence Platform
Education
MSc Computer ScienceReykjavík University, 2017 – 2021. Artificial Intelligence and Data Science. Thesis: Recommendation Systems.
BSc MathematicsUniversity of Iceland, 2011 – 2016. Computational Mathematics and Computer Science.
Materials
AI & agentspydantic-ai, MCP, Anthropic/OpenAI/Gemini APIs, agent evaluation, RAG and embeddings, durable agent workflows (DBOS), multi-agent engineering teams, local inference, video diffusion models
LanguagesPython (expert), SQL, TypeScript, Go, R, Logica
Backend & APIsFastAPI, Pydantic, OpenAPI, typed client generation, OAuth 2.0 (RFC 7591/8414/9728), event sourcing, idempotent and replayable APIs
DataPostgres, DuckDB/DuckLake, Snowflake, Neo4j, Microsoft Fabric/PySpark, Kimball modelling, Power BI, forecasting
InfrastructureNixOS, Docker, Azure (Container Apps, Bicep), AWS, Cloudflare, Tailscale, RunPod, GitHub Actions with self-hosted runners
Engineering & observabilityDomain-driven design, architecture enforced in CI (import-linter, tach), typed error hierarchies, Logfire/OpenTelemetry, uv/ruff/ty
Project
Hákon Freyr Gunnarsson
Sheet
6 / 6
Title
Experience
Scale
N.T.S.
Rev
2026-10