Ярослав

Yaroslav Vasylenko
AI Behavior Architect · LLM Systems Engineer · Prompt Architecture · Agent Orchestration

GitHub
github.com/neuron7x
Telegram
@neuron7x
Viber
[открыть контакты](см. выше в блоке «контактная информация»)
Location
Ukraine · Remote

PROFILE
AI/LLM systems engineer focused on instruction architecture, agent orchestration, and LLM behavior evaluation. I design system prompts, instruction hierarchies, role contracts, routing logic, guardrails, tool-using workflows, and evaluation protocols. I treat AI systems as software: explicit invariants, schema-bound outputs, typed contracts, tests, CI gates, reproducible execution, and evidence artifacts.

CORE COMPETENCIES

Prompt Architecture
Agent Orchestration
Evaluation & Reliability
Python / Delivery
– System prompts, role specifications
– Instruction hierarchy, scope boundaries
– Decision logic, negative constraints
– Behavior contracts for stable model output
– Tool-using agents, multi-step workflows
– Model / role / tool routing
– Context boundaries, handoff logic
– Human-in-the-loop review contours
– Failure-mode analysis, root-cause review
– Hallucination control, verification gates
– Schema-bound outputs, deterministic contracts
– Benchmark and protocol design for LLM behavior
– CLI / API surfaces for AI workflows
– Tests, CI gates, type checks, manifests
– Health / readiness probes, logs, observability
– Reproducible execution, evidence bundles

ADVERSARIAL ORCHESTRATION METHODOLOGY
Production AI systems built on a multi-node adversarial pipeline. Each downstream node independently verifies the assumptions of the upstream node — it does not continue its output. Execution chain: Creator → Critic → Auditor → Verifier → Test Gate → CI. Evidence gaps, schema violations, and ambiguity are not suppressed — outputs are downgraded or blocked at the gate where they fail.

• SCL Protocol (Signal–Compression–Lock) — 6-gate prompt verification standard applied before every inference call.
• Fail-closed gates — no output passes without explicit verification at each stage boundary.
• Evidence Confidence tiering — outputs carry confidence levels; insufficient evidence triggers rejection, not hallucinated fill.
• Anti-gaming enforcement — audit nodes operate from independent context, not inherited from previous node output.

INDEPENDENT PRACTICE
2023 — present

• Designed and delivered 6 audit-grade AI security evaluation protocols with fail-closed gates, Evidence Confidence tiering, and anti-gaming enforcement. Verified zero prior art for the specific protocol combination. 95 passing pytest tests. MANIFEST integrity verification. Dual licensing: open-source (AGCL-1.0) + commercial.
• Built a typed intermediate representation for LLM inference (Reasoning Contract Grammar), modeled on LLVM IR: ~5,700 LOC, 443 tests, 90% branch coverage, mypy strict 0 errors. Academic paper targeting NeurIPS/ICML workshop track.
• Produced 5 specialized prompt instruments for a retail intelligence platform (tenetix.ai): behavioral specification, routing rules, negative constraints, schema-bound output contracts. Delivered as production-ready instruction packages.
• Implemented a neurobiologically-grounded LLM wrapper system: 9,723 LOC, 421 test functions, 22 CI/CD workflows, formal TLA+/Coq specifications. Modules: MoralFilterV2, CognitiveRhythm, AphasiaBrocaDetector.
• Built a production React/Vite + Python backend + Cloudflare Worker BFF system (AXL-AD2026): PRODUCTION_SPEC_V2.1, 9 hard gates passing, SSDF/SLSA compliance, reproducible builds, proof bundles.
• Formalized and published a 4-Node Web Orchestration methodology (IDEA → AGENT·2 → AUDIT → CODEX → CI → LOCAL → MERGE): complete system instructions, context-passing rules, failure mode map FM-01–FM-07, weekly optimization loop protocol.

EXECUTION LOGIC
• Specification precedes generation — behavioral spec, invariants, and failure modes are defined before any LLM call.
• Each subsequent node verifies the assumptions of the previous one — it does not continue its output.
• Lack of evidence, ambiguity, or schema violation is not masked — the output is downgraded or blocked.
• Prompts and protocols are versioned engineering artifacts: routing rules, failure modes, tests, and controlled output.

TECHNICAL STACK

Python
LLM Systems
Prompt Architecture
AI Agents
Evaluation Frameworks

CLI / API
Schema Validation
CI / Testing
Reproducibility
Observability

PUBLIC WORK
GitHub: github.com/neuron7x — Production AI evaluation protocols, typed LLM inference IR, agent orchestration systems. All repositories include CI/CD pipelines, test coverage reports, strict type checking, and reproducible execution instructions.

DIRECT NOTE
My professional background is entirely independent — no corporate employment history. Every system listed above was designed, built, tested, and shipped solo: no team scaffolding, no inherited architecture, no fallback reviews. Every architectural decision and every production failure was mine to diagnose and fix. I am applying to work within a team because the problems worth solving at scale require shared infrastructure — not because I need institutional support to build.

Yaroslav Vasylenko · github.com/neuron7x · @neuron7x · Ukraine · Remote