Ivan Romanteyev

ML/AI-engineer.

[Ivan Romanteyev @IvanRomanteev]

I BUILT TRUSTFLOW — AN AI SYSTEM THAT DOESN'T JUST ACT. IT PROVES WHAT HAPPENED.

After an intense engineering sprint, I'm excited to officially launch TrustFlow for the Meta Global AI Developer Hackathon 2026! 🏆

Here's the problem I wanted to solve:

What happens when an autonomous AI agent executes a financial operation, times out, and nobody knows whether the transaction actually happened?

Do you retry it?

What if the money was already transferred?

What if the agent confidently reports success without sufficient evidence?

This is where AI agents stop being just impressive demos — and start needing real engineering.

! INTRODUCING TRUSTFLOW !

An experimental framework for verifiable AI-agent workflows, built around one principle:

AI can propose. Humans authorize. Evidence proves.

Instead of blindly trusting an LLM's decisions, TrustFlow separates probabilistic reasoning from deterministic execution controls.

⚙️ WHAT'S UNDER THE HOOD?

🔹 Multi-step AI agent orchestration

🔹 Structured planning powered by Meta Muse

🔹 Deterministic policy and risk enforcement

🔹 Human-in-the-loop approval bound to exact SHA-256 plan hashes

🔹 Idempotent operations to prevent duplicate execution

🔹 Independent ledger verification

🔹 Recovery after timeouts and process crashes

🔹 Persistent workflow state, audit trails, and evidence tracking

🔹 Automated evaluation, regression testing, and browser-based verification

🤖 BUILT WITH META. POWERED BY META.

Meta Muse Code was a major AI-assisted engineering collaborator throughout development — helping with architecture, implementation, testing, debugging, and reliability improvements.

I also used OpenAI Codex for implementation work and independent engineering reviews.

Built TrustFlow through an intensive Muse Code-assisted engineering workflow. Our Meta Model API dashboard recorded 322M+ processed tokens and ~3.5K requests during the development period.

At runtime, TrustFlow integrates the Meta Model API with:

⚡ Muse Spark 1.3 — structured agent planning

🧠 Muse Wonder 1.0 — semantic review

📊 REAL TEST RESULTS

✅ 10 real Muse Spark API calls

✅ 1 real Muse Wonder semantic-review call

✅ Spark test matrix: 5/5

✅ 654 automated tests passed, 19 skipped

✅ Independent ledger checks and controlled crash-recovery demonstrations

The Wonder review used an offline-prepared synthetic scenario; full Wonder/mixed-model validation is not claimed.

All financial operations use synthetic funds. This is an engineering prototype, not a production banking platform.

💡 WHY THIS MATTERS

As AI agents move from answering questions to executing real-world actions, reasoning quality alone won't be enough.

We need authorization boundaries.

We need stable operation identities.

We need independent evidence.

And we need systems that can recover from failure without blindly repeating irreversible actions.

That's what TrustFlow is exploring.

🎬 WATCH THE FULL 3-MINUTE DEMO:
https://t.co/OGOPp7L87S

💻 OPEN-SOURCE GITHUB:
https://t.co/JYmlWMyZ84

👨‍💻 MY GITHUB:
https://t.co/4RFdrw4ixk
💻 MY LINKEDIN:
https://t.co/Mpxym3xYI2
❤️ If you find this project interesting, I'd genuinely appreciate a like, repost, or comment.

I'd especially love feedback from people working on AI agents, LLM infrastructure, distributed systems, and production AI.
What do you think is the biggest unsolved reliability problem in autonomous AI agents? 👇

#MetaAI #MetaHackathon #MuseCode #MuseSpark #MuseWonder #OpenAI #Codex #AI #ArtificialIntelligence #MachineLearning #DeepLearning #GenerativeAI #AgenticAI #AIAgents #LLM #MultiAgentSystems #Python #MLOps #LLMOps #AIEngineering #FinTech

[Ivan Romanteyev @IvanRomanteev]

Great question! In the current TrustFlow prototype, the transfer and ledger write are part of the same atomic SQLite transaction (BEGIN IMMEDIATE -> COMMIT). So there’s no separate external payment that can succeed before our ledger records it.

If a timeout occurs, the operation becomes EXECUTION_UNCERTAIN. Recovery checks the original idempotency key against the persisted ledger rather than trusting the plan hash or blindly retrying the transfer.

For a real payment provider, we'd need to reconcile against its authoritative transaction status using the same stable operation ID/idempotency key. The plan hash proves what was authorized, not what actually executed.

That external reconciliation layer is a future integration, not something the current prototype claims to implement.

[Ivan Romanteyev @IvanRomanteev]

I BUILT TRUSTFLOW — AN AI SYSTEM THAT DOESN'T JUST ACT. IT PROVES WHAT HAPPENED.

After an intense engineering sprint, I'm excited to officially launch TrustFlow for the Meta Global AI Developer Hackathon 2026! 🏆

Here's the problem I wanted to solve:

What happens when an autonomous AI agent executes a financial operation, times out, and nobody knows whether the transaction actually happened?

Do you retry it?

What if the money was already transferred?

What if the agent confidently reports success without sufficient evidence?

This is where AI agents stop being just impressive demos — and start needing real engineering.

! INTRODUCING TRUSTFLOW !

An experimental framework for verifiable AI-agent workflows, built around one principle:

AI can propose. Humans authorize. Evidence proves.

Instead of blindly trusting an LLM's decisions, TrustFlow separates probabilistic reasoning from deterministic execution controls.

⚙️ WHAT'S UNDER THE HOOD?

🔹 Multi-step AI agent orchestration

🔹 Structured planning powered by Meta Muse

🔹 Deterministic policy and risk enforcement

🔹 Human-in-the-loop approval bound to exact SHA-256 plan hashes

🔹 Idempotent operations to prevent duplicate execution

🔹 Independent ledger verification

🔹 Recovery after timeouts and process crashes

🔹 Persistent workflow state, audit trails, and evidence tracking

🔹 Automated evaluation, regression testing, and browser-based verification

🤖 BUILT WITH META. POWERED BY META.

Meta Muse Code was a major AI-assisted engineering collaborator throughout development — helping with architecture, implementation, testing, debugging, and reliability improvements.

I also used OpenAI Codex for implementation work and independent engineering reviews.

Built TrustFlow through an intensive Muse Code-assisted engineering workflow. Our Meta Model API dashboard recorded 322M+ processed tokens and ~3.5K requests during the development period.

At runtime, TrustFlow integrates the Meta Model API with:

⚡ Muse Spark 1.3 — structured agent planning

🧠 Muse Wonder 1.0 — semantic review

📊 REAL TEST RESULTS

✅ 10 real Muse Spark API calls

✅ 1 real Muse Wonder semantic-review call

✅ Spark test matrix: 5/5

✅ 654 automated tests passed, 19 skipped

✅ Independent ledger checks and controlled crash-recovery demonstrations

The Wonder review used an offline-prepared synthetic scenario; full Wonder/mixed-model validation is not claimed.

All financial operations use synthetic funds. This is an engineering prototype, not a production banking platform.

💡 WHY THIS MATTERS

As AI agents move from answering questions to executing real-world actions, reasoning quality alone won't be enough.

We need authorization boundaries.

We need stable operation identities.

We need independent evidence.

And we need systems that can recover from failure without blindly repeating irreversible actions.

That's what TrustFlow is exploring.

🎬 WATCH THE FULL 3-MINUTE DEMO:
https://t.co/OGOPp7L87S

💻 OPEN-SOURCE GITHUB:
https://t.co/JYmlWMyZ84

👨‍💻 MY GITHUB:
https://t.co/4RFdrw4ixk
💻 MY LINKEDIN:
https://t.co/Mpxym3xYI2
❤️ If you find this project interesting, I'd genuinely appreciate a like, repost, or comment.

I'd especially love feedback from people working on AI agents, LLM infrastructure, distributed systems, and production AI.
What do you think is the biggest unsolved reliability problem in autonomous AI agents? 👇

#MetaAI #MetaHackathon #MuseCode #MuseSpark #MuseWonder #OpenAI #Codex #AI #ArtificialIntelligence #MachineLearning #DeepLearning #GenerativeAI #AgenticAI #AIAgents #LLM #MultiAgentSystems #Python #MLOps #LLMOps #AIEngineering #FinTech

[Ivan Romanteyev @IvanRomanteev]

What happens when an AI agent executes a transaction, crashes right after COMMIT, and never receives confirmation?

A blind retry could execute the same operation twice.

That's exactly what I built TrustFlow to address.

665 automated tests passed
18 Docker tests passed
3 complete Chromium demo replays
Real Meta Muse Spark & Wonder integration
TrustFlow demonstrates independently verifiable execution and recovery without duplicating synthetic transactions.
Built for the Meta Global AI Developer Hackathon 2026.
I'd love your feedback! If you find this useful, a repost would help bring the project to more AI engineers. 🤝

[Ivan Romanteyev @IvanRomanteev]

Great question. In TrustFlow, approval is bound to the entire immutable plan snapshot, including its version and SHA-256 hash — not to individual steps.

If a step changes before execution, the revised plan has a different hash, so the previous approval cannot authorize it. A new high-risk plan requires fresh human approval of the full snapshot. Per-step approvals aren't supported in the current prototype.

Importantly, an execution timeout doesn't trigger blind re-planning or re-execution. TrustFlow reconciles the original operation ID against the ledger and performs verification-only recovery.

This prevents a changed plan from silently inheriting authorization or an uncertain transfer from being executed twice.

[Ivan Romanteyev @IvanRomanteev]

Exactly. That's why TrustFlow keeps the critical verification deterministic rather than relying on another LLM to confirm success.

The verifier checks the approved plan hash, operation identity, committed SQLite transaction, payload consistency, and ledger conservation. These are relatively inexpensive checks against persisted state.

Semantic review is a separate layer and cannot override a deterministic failure.

The current prototype doesn't provide cryptographically signed or tamper-proof proofs yet, but it demonstrates how to verify execution without trusting the agent's own summary.

[Ivan Romanteyev @IvanRomanteev]

@Dedaj371505 good luck on the hackathon bro