Applied AI systems architecture
Provider-neutral worker boundaries, local sovereignty, modular control planes, and explicit state-to-outcome loops.
Andriy Baygerych / professional profile
Founder-operator and applied AI systems engineer.
I design, operate, audit, and progressively govern machine systems under real constraints. My strongest work is at the boundary between model capability and host-owned reality: accepted context, tools, writable surfaces, independent verification, retained failure memory, and authority.
Professional summary
I make complex AI systems answerable to evidence.
Independent founder-operator of IFP/SAGE, a local-first governed evidence and learning substrate. I coordinate high-compute AI workers for repository archaeology, diagnosis, implementation, and adversarial review while preserving a host-owned separation between worker output, verified facts, retained learning, resource use, and authority.
Non-traditional path. I do not claim a formal IT engineering diploma. I present live systems, exact artifacts, bounded results, and the defects found in my own work.
Core capabilities
Live-system re-derivation
Runtime identity, service topology, writer/reader/consumer graphs, stale-document correction, and evidence status.
AI evals and closure
Distinguishing execution, use, changed later behavior, verified improvement, and cross-provider transfer.
Governance engineering
Admission boundaries, capability envelopes, authority containment, rollback, refusal, and append-only correction.
Technical programme direction
Scope locks, stop conditions, worker-role separation, artifact review, independent challenge, and staged authorization.
Machine-readable truth surfaces
Human and machine documentation, claim-state vocabularies, structured data, public verification paths, and honest nulls.
Selected work
Machine Governor P0-RL
Re-derived 35 learning/control mechanisms, separated internal learning from worker-steering learning, and reduced the missing delta to a deterministic steering consumer over Track-2, G1, and KAS. Primary result provisional; artifact review pending.
DecisionBundleV1 P1 evidence rebuild
Qualified an exact-SHA candidate with 315 passing clean-worktree tests. The verifier established package internal consistency and refused the unsupported broader-record claim.
OVL paid attestation rail
Closed a self-paid Base Sepolia testnet issuance with x402 settlement, Ed25519 signing, independent content-hash re-derivation, hash-chain correlation, and four tamper rejections.
Track-2 generalized closure ledger
Directed the reduction of agent invocation, governance admission, closure event, and substrate consumption into append-only records with runtime hooks and explicit authority separation.
Governed knowledge and experiment isolation
Developed the architectural path for admitting machine proposals, holding them in non-authoritative lanes, replay-checking them, and isolating experimentation from production learning.
Operating principles
- A worker's completion report is a claim, not programme closure.
- Historical documents are evidence inputs, not current runtime truth.
- A stored memory is not learning until it changes a later attempt.
- Changed behavior is not improvement until a held-out result verifies it.
- Improvement does not grant authority.
- Negative and undecidable results remain first-class records.
- The best architecture is often the smallest one that closes the actual loop.
Role fit
Where I am most useful.
Forward-deployed AI engineering
Embedding advanced models into messy customer or operator environments and carrying the work through to a checkable outcome.
Agent systems and infrastructure
Context, tools, capability boundaries, memory, retries, verification, closure, and provider-neutral control.
AI evaluations and reliability
Adversarial testing, failure attribution, honest baselines, tamper detection, and downstream-effect verification.
Applied AI architecture and technical product
Turning a difficult operational problem into a bounded system, evidence model, delivery plan, and operator-visible surface.