🚧 Challenges I Ran Into

Building COMPUT3 involved solving problems across systems, security, and Web3, and one of the biggest challenges was ensuring true data confidentiality in a hostile environment.

---

🔐 1. Preventing Provider Access to Container Data

A major hurdle was:

How do you run a container on someone else’s machine without letting them access your code or data?

By default, even with Docker isolation, a host-level operator can:

• Inspect volumes
• Read filesystem data
• Attach debuggers

✅ Solution

We implemented disk-level encryption using LUKS at the workspace level:

• Each container is backed by an encrypted volume
• The encryption key is not stored, but derived dynamically from on-chain attestations
• Even if someone copies the entire disk, it remains unreadable ciphertext

We also ensured:

• Keys are only available during mount time
• Temporary key files are immediately wiped
• Access can be revoked by invalidating the attestation

This shifted the model from:

“trust the operator” → “make access cryptographically impossible”

---

⚙️ 2. Verifiable Agent Execution

Another challenge was:

How do we prove that the AI agent didn’t modify or inject anything during deployment?

AI systems are inherently non-deterministic, so we needed a way to make execution auditable and tamper-evident.

✅ Solution

We built a structured logging system:

• Every agent action (tool calls, container ops, installs) is:
  • Logged
  • Hashed
• All hashes are aggregated into a Merkle tree
• The root is attested on-chain via EAS

This allows users to:

• Recompute the hash
• Verify execution independently

---

💰 3. Designing Trustless Payments

We initially implemented simple request-based payments, but ran into a key issue:

What if a provider takes payment and then shuts down the service?

✅ Solution

We redesigned the system into a hybrid payment model:

• Partial upfront payment for deployment
• Remaining payments streamed based on uptime
• Providers must continuously submit proofs
• Payments stop immediately if the service stops
• Malicious providers can be slashed via stake

---

🧠 Key Learning

The biggest takeaway was:

You can’t rely on trust assumptions — you have to redesign systems so that trust is unnecessary.

This meant combining:

• Cryptography
• System isolation
• On-chain verification
• Incentive design

into a single cohesive architecture.

---

💣 Final Reflection

Building COMPUT3 wasn’t just about deploying apps — it was about rethinking how infrastructure works when trust is removed from the system.