AI-Native Systems Architect · Creator of EvidenceBound · Founder of SignalReview
EvidenceBound · ruslan@evidencebound.org · LinkedIn · ORCID
I design agentic AI systems that fail closed when evidence, dependencies, integrity, or policy state can no longer be trusted.
My work spans deterministic verification, bounded execution, tamper-evident Proof Packs, verifiable memory, trust-graph recovery, multi-agent orchestration, security boundaries, CI/CD, and production acceptance.
Open to research/consortium collaboration around trustworthy agentic systems and to remote AI systems architecture, applied AI, agent platform, verification, and technical product engineering roles.
Institutional site: https://evidencebound.org
Canonical OSS core: https://github.com/moneyparking/evidencebound-core
Project contact: ruslan@evidencebound.org
EvidenceBound is a systems approach for making AI-generated actions inspectable, reproducible, challengeable, and blockable.
Current evidence
↓
Deterministic verification
↓
Tamper-evident proof
↓
Verifiable memory
↓
Trust-break detection
↓
Selective recovery
↓
Mandatory human or policy-controlled decision
The LLM may interpret bounded evidence. It does not grant itself trust, hide unavailable inputs, override deterministic gates, or authorize unsafe side effects.
| Project | Architectural problem | Verified implementation |
|---|---|---|
| EvidenceBound Core | Bind agent state to evidence/provenance/policy and recover selectively after trust changes | Framework-agnostic Python runtime, deterministic verification, dependency blast-radius analysis, fail-closed selective recovery, signed-receipt and persistence seams, conformance corpus, supply-chain attestations |
| EvidenceBound Recovery Mesh | Recover an agent fleet after one checkpoint becomes untrustworthy without restarting safe work | Google ADK, Vertex AI / Gemini 3.5 Flash, Cloud Run, Agent Registry, Secret Manager, Workload Identity Federation, Trust Graph blast-radius analysis, fail-closed action gate, selective recomputation |
| EvidenceBound DataHub Gate | Decide whether an agent-generated data action is bound to current metadata before human review | Official DataHub MCP, schema and bounded-lineage binding, Restricted AST Policy, no-exec bounded interpreter, SHA-256 Proof Packs, Ed25519 seals, native DataHub description write-back |
| EvidenceBound Verified Memory | Reopen prior agent state and prove what changed rather than merely recalling an answer | CockroachDB persistence, canonical snapshots, integrity verification, T0 → T1 comparison, AWS Lambda, S3, CloudFormation, GitHub OIDC → AWS STS |
| Governed Codex Module Factory | Generate analytics modules without granting generated code automatic deployment authority | Codex workflow, isolated execution, validation gates, sealed review artifacts, mandatory human review |
| Forensic Multi-Agent Sports Intelligence | Produce a bounded verdict from adversarial agent roles without inventing unavailable provider facts | Qwen Cloud, four-role evidence-first review, visible missing-data states, orchestrated bounded verdict |
| SignalReview | Turn structured match evidence into a retained, transparent decision-support workflow | Next.js, TypeScript, React, Python, FastAPI, PostgreSQL, Supabase, Docker, GitHub Actions, Vercel, Render, Cloudflare |
A controlled production run compared a full restart with selective recovery:
- full restart: 4 model calls / 1,781 input tokens;
- selective recovery: 3 model calls / 1,358 input tokens;
- saved in that exact run: 1 model call / 423 input tokens (~24%).
This is a measured result from one controlled run, not a universal savings claim. The current bounded Cloud Run slice does not claim durable multi-week memory.
git clone https://github.com/moneyparking/evidencebound-datahub-gate.git
cd evidencebound-datahub-gate
make test-reproExpected retained states include:
VERIFIED
BLOCKED: SCHEMA_MISMATCH
REPRODUCED
ARTIFACT_TAMPERING_DETECTED
SIGNATURE_VALID
VERIFIED is deliberately narrow. It does not mean production authorization, data truth, certification, or permission to deploy.
- AI-native systems and agent platform architecture
- Agentic AI, MCP, Google ADK, OpenAI, and Qwen Cloud
- Deterministic runtimes, restricted execution, and fail-closed controls
- Verification contracts, canonicalization, provenance, and content addressing
- Verifiable memory, dependency graphs, invalidation, and selective recovery
- Python, TypeScript, Next.js, React, FastAPI, Pydantic, SQL
- PostgreSQL, Supabase, CockroachDB, AWS, Google Cloud, Vercel, Render, Cloudflare
- Docker, GitHub Actions, OIDC/WIF, pytest, Ruff, Mypy, security and release gates
- Product strategy, monetization, QA, deployment, and production acceptance
- EvidenceBound institutional site
- EvidenceBound Core
- EvidenceBound maintainer profile
- Machine-readable EvidenceBound identity
- EvidenceBound reference implementations
- EvidenceBound research lineage
- ORCID
- Historical EvidenceBound-MAS whitepaper
- Recovery Mesh on Devpost
- DataHub Gate on Devpost
- Verified Memory on Devpost
For EvidenceBound research, OSS, security and consortium work: ruslan@evidencebound.org.
Public materials intentionally exclude private repositories, credentials, customer information, personal address, telephone number, and transient location.