Skip to audit workspace
Sealed Ledger
Start audit

NEAR Hackathon 2026 · Track 03

Know where your money went. Without telling anyone.

Sealed Ledger reads receipts and bank statements inside NEAR AI Cloud's confidential TEE, categorizes every charge, finds the anomalies, and proves the analysis stayed private. No account. No storage. No plaintext outside the enclave.

Run a private auditSee the proof chain
Intel TDXNVIDIA NRAS0 bytes retainednear-ai-cloud skill

Audit certificate

TEE verified

Verified

Hardware proof attached

Intel TDX quote + NVIDIA GPU evidence

10/10
Model
Qwen/Qwen3.8-27B
CPU TEE
Intel TDX · dcap-qvl
GPU TEE
NVIDIA NRAS · verified
Signer
0x71c4…4e2 · bound
Retention
0 bytes
nonce8f3a91c4e7b2…c21b4e7a90aa
Fresh per auditproof.json

Private workspace

Run an audit.

Upload evidence, watch the enclave work, then inspect the proof. Your files are held in memory for one request and never stored.

Open focused workspace
API key required0 bytes retained

01 · Evidence

Upload receipts & statements

Receipt photos or a bank CSV. Parsed in memory, never written to disk.

In-memory

Drop files here, or browse

PNG, JPG, WebP receipts · CSV statement

6 images · 12 MB · 0 bytes stored

No data? Use the fictional sample.

A 24-row September statement plus five receipts, seeded with duplicates, a subscription jump, and a late-night charge.

Server key missing. Add NEARAI_CLOUD_API_KEY to .env.local and restart npm run dev.

TEE model selected at runtime

Add a receipt or statement to run.

02 · Proof

Live verification

Every step leaves a check in the certificate.

TDX
Waiting for evidence0/6
  1. 01Read filesParse the statement and receipts in server memory
  2. 02Select TEE modelConfirm the model is NEAR AI-hosted and attestable
  3. 03Attest hardwareFetch a fresh nonce-bound report for model and gateway TEEs
  4. 04Analyze privatelyCategorize and flag anomalies inside the enclave
  5. 05Verify signatureMatch request/response hashes to the TEE signer
  6. 06Check TDX + GPUValidate Intel TDX quotes and NVIDIA NRAS evidence

Proof stack

256-bit noncedcap-qvlNVIDIA NRASECDSA recovery

How it works

A private auditor with a paper trail.

Three moves: evidence in, analysis inside the enclave, proof out. Every claim in the result is backed by a check you can inspect.

01

Evidence in

Drop receipt photos or a CSV statement. The server parses them in memory and prepares one private request.

02

Analysis inside the enclave

A NEAR AI-hosted model categorizes every row and flags anomalies inside an Intel TDX VM with an NVIDIA TEE GPU.

03

Proof out

The app verifies the TDX quote, GPU evidence, response hashes, and ECDSA signer before showing any result.

The trust chain

Hardware first.
Signature second.

A signature alone proves someone signed bytes. Binding that signature to genuine Intel TDX and NVIDIA hardware is what makes the result verifiable.

Chain of trust

NVIDIA GPU → Intel TDX quote → signing address → chat signature → exact request + response bytes

Checklist

What the Verified panel proves

10 checks
  • TEE-hosted model selected from /v1/models
  • Fresh 256-bit nonce per request
  • Intel TDX quote validated with dcap-qvl
  • TDX report data binds the signing address + nonce
  • NVIDIA NRAS overall and per-GPU verdicts
  • GPU payload nonce matches the request nonce
  • Compose manifest compared with mr_config
  • Request + response SHA-256 match the signature
  • ECDSA signer recovers to the attested TEE address
  • No database, disk writes, content logs, or cache

Privacy design

The best retention policy is having nothing to retain.

The server processes your files, then forgets them. The only external destination is the NEAR AI Cloud TEE.

Leaves your browser

Only the files you choose, over TLS to cloud-api.near.ai

Server memory

Parsed once, held only until this response closes

Written to disk

Nothing · no database, no logs, no cache

Seen by the host

Ciphertext and attestation metadata, never prompt plaintext

No-retention in code: the route handler holds uploads in a local array, passes them to the TEE request, returns the result, and ends. There is no database client, no filesystem write, no content logger, and no response cache.

Built for Track 03

Privacy & Private AI, end to end

  • OpenAI-compatible NEAR AI Cloud API at cloud-api.near.ai/v1
  • TEE-hosted model chosen from /v1/models
  • Fresh-nonce attestation rendered as a Verified panel
  • Zero server-side retention, documented in code and README
  • Fictional samples so judges never need real data
  • Minimal verification script before the UI, using the near-ai-cloud skill
NEAR AI Cloud quickstart