Andriy Baygerych / founder-operator / applied AI systems

I build machine systems that have to prove what happened next.

I am an independent applied AI systems engineer in Norway and the founder-operator of IFP/SAGE. My work sits between live-system archaeology, agent infrastructure, independent verification, authority containment, and forward-deployed problem solving. I use AI workers aggressively, but I do not let a worker certify its own result.

Local workerQwen / target
External workerGrok / no authority
Visual worldEVOKE / benchmark
AuthoritySAGE / unchanged
Current programmeSAGE_MACHINE_GOVERNOR_001
P0-RLPROVISIONAL / REVIEW HELD
Compute integrationISOLATED BENCHMARK ONLY
Open toROLES / CONTRACTS / AUDITS

Current P0-RL provisional findings

35learning and control mechanisms mapped
21with current firing evidence or current output
0verified L2-L4 external-worker improvement loops
1thin deterministic steering component identified

P0-RL primary finding is provisionally accepted. Artifact review and bounded Codex challenge remain pending as of 24 August 2026.

The work, not the title

No conventional IT diploma. A live system is the credential.

I did not enter software through a standard computer-science career path. I built and operate a complex local AI substrate, direct multiple machine workers, and keep negative results in the record when the evidence does not support the original idea.

Read the full profile

01 / RE-DERIVE

Recover reality from live systems

I distinguish current runtime behavior from stale architecture documents, class names, worker self-reports, and optimistic dashboards.

02 / GOVERN

Keep machine workers inside host-owned boundaries

Task contracts, accepted context, tools, writable paths, verification depth, closure, and authority remain properties of the host system.

03 / VERIFY

Separate activity from demonstrated improvement

A model answer, passing unit test, stored memory, or completed task report is not treated as proof that a downstream result improved.

04 / REDUCE

Find the smallest correct intervention

The Machine Governor audit rejected a new learning database and new general engine. The missing delta reduced to one deterministic steering consumer.

What I can do for a team

Forward-deployed judgment for agentic systems.

I am strongest where a powerful model meets a real operating environment: incomplete context, ambiguous authority, unreliable self-report, legacy behavior, downstream dependencies, and a customer or operator who needs a result that can survive examination.

Live-system re-derivation

Trace writers, readers, consumers, services, state, and runtime identity before accepting architectural claims.

Agent closure and evals

Define whether an invocation merely ran, was consumed, changed a later attempt, improved a held-out result, or transferred across providers.

Authority and containment

Keep model output, evidence, resource use, value, and authority structurally separate.

Provider-neutral worker integration

Design host-owned task, context, capability, verification, rollback, and closure contracts around replaceable models.

Evidence-grade public surfaces

Expose what is proven, what is attributed, what is provisional, and what remains impossible to verify.

Technical programme leadership

Coordinate specialized AI workers, preserve stop conditions, adjudicate artifacts, and prevent unfinished loops from being called done.

Operating doctrine

Instructions steer the current attempt. Verified closure memory should steer the next attempt.

SAGE already learns internally. The active programme asks a narrower and harder question: can independently verified consequences change how the next unfamiliar machine is tasked, constrained, checked, and trusted?

Agreed compute architecture

Four roles. One authority boundary.

The external model may be more capable. The visual model may generate worlds. The local model may work offline. None of them inherits state authority from capability.

01 / LOCAL

Qwen 27B target

Primary local and offline language worker. Exact package identity and runtime envelope must be verified before installation or routing.

Identity check required
02 / EXTERNAL

Grok

High-compute reasoning, research, architecture and coding. Its outputs enter through provenance, admission, verification and closure.

Authority: none
03 / VISUAL

EVOKE

Candidate local visual-world and simulation worker. Installation is isolated; residency, speed, thermals and licensing require measurement.

Benchmark pending
04 / GOVERNOR

SAGE

Sole owner of accepted state, decisions, admission, independent verification, closure, retained steering and revocation.

Authority unchanged

IFP / SAGE

The system is the proof environment.

IFP is the wider platform direction. SAGE is the live local substrate: state, learning, admission, evidence, regulation, operator truth, and current Machine Governor work. External models are replaceable workers; SAGE is intended to remain the owner of state, closure, and authority.

External workerGrok / Qwen / Codex
WorkerDoortask + context + capability
Resultcandidate / action / artifact
Verificationindependent of worker
Track-2 + G1 + KASfacts + shadow + class
steering_updatethin deterministic consumer
Next invocationmaterially changed envelope

Live operator surface

The substrate remains visible.

The embedded surface is served from infinity-folder.org. It is a separate live system surface, not a static claim generated by this portfolio site.

https://infinity-folder.org/index.html
Open independently
Embedded cross-domain content can be blocked by the live site's frame policy or by a browser extension. The independent link above is the authoritative fallback.

Work with me

Roles, bounded engineering work, and serious technical conversations.

I am open to forward-deployed AI, agent systems, AI evaluations and reliability, applied AI architecture, technical product work, and focused consulting where the output must be independently checkable.

A bounded first engagement

Agentic AI closure and steering audit: inventory claimed memory and feedback paths, classify L0-L4 closure, locate where history fails to change later behavior, identify authority leakage, and propose the smallest deterministic correction with held-out verification.

  • agent infrastructure
  • evals
  • failure attribution
  • authority
  • worker steering
  • evidence package