XIO Compliance Brain
Triad Review Engine for audit-ready compliance work.
Upload a compliance document. Three AI reviewer perspectives — Counsel, Risk, and Evidence — find gaps, verify citations, expose disagreement, and create approval-ready work product.
Judges should start with the seeded demo. The live debate is optional and depends on the AMD droplet being warm.
Latest Triad Review · seeded sample
open →Ontario OM Compliance Review — North Capital Series A
OM compliance review · Ontario · NI 45-106
- Compliance score
- 62%
- Critical gaps
- 3
- Verified citations
- 5
- Needs verification
- 2
- Reviewer disagreement
- 1 material
- Export
- Blocked · pending approval
Start here for judging
A clean 90-second path through the demo
1. Start
Open the seeded review
No upload, no GPU wait, always available.
2. Inspect
Compare one answer vs Triad
See gaps, citations, disagreement, and blocked export.
3. Verify
Expand the audit chain
Hash-linked rows show who did what and which provider served it.
4. Optional
Run the live debate
Use this only if the AMD endpoint is warm.
The triad
Regulatory Counsel
Rule breaches · missing disclosures · jurisdiction issues
Finds rule breaches, missing disclosures, and jurisdiction issues against the applicable corpus (NI 45-106, OSC Rule 45-501, NI 31-103).
Risk Officer
Severity · investor exposure · operational risk
Scores business impact, severity, and operational exposure. Flags issues that would draw a regulator's eye even when technically defensible.
Evidence Auditor
Citation verification · stale authority · source gaps
Checks citations, stale authorities, unsupported claims, and source gaps. Refuses to sign off on findings without a verifiable authority.
Use cases
Offering memorandum review
NI 45-106 / OSC Rule 45-501
KYC / AML file review
NI 31-103 Part 13 · FINTRAC
Marketing compliance review
NI 81-102 · prohibited representations
Regulator letter response
OSC / CIRO / FINTRAC inquiries
Privacy / data handling memo
PIPEDA · Quebec Law 25
Contract clause risk review
Liability · indemnity · termination
Same Triad Review pipeline, different corpora and reviewer prompts per task type.
What it costs · what it saves
Lawyer-hours per OM review
3–5 hrs
Industry baseline: a junior associate first-pass on a private-placement OM.
Inference cost per Triad Review (3 voices + synthesis)
≈ $0.04
Measured on this app: ~6,500 output tokens × $0.0006/1K (self-hosted MI300X amortised cost).
Cloud equivalent for the same workload
≈ $8.10
3 voices × 32K-ctx 70B-class on a hosted API at $1.20/1M output. Self-hosting flips a 200× ratio.
Numbers are from this codebase's synthetic test runs against the live AMD droplet — not a customer engagement. Real-world spread varies with deal complexity. The argument is order-of-magnitude: GPU memory headroom unlocks ensembles that cloud APIs price out of reach.
Why AMD matters
Triad Review on AMD Instinct MI300X
Model
Qwen 2.5 72B-Instruct
32K context · FP16
Serving
vLLM (OpenAI-compatible)
Streaming SSE · /v1/*
Hardware target
AMD Instinct MI300X
192 GB HBM3
Workflow
3 voices + synthesis
Optional Round 2 (defend / update / concede)
Large-memory GPU serving fits a 72B model and a three-voice ensemble on a single card. The same workload on cloud APIs requires roughly 4× H100s — GPU memory headroom is what makes Triad Review economically possible.
The hosted preview at compliance-ai-amd-demo-production.up.railway.app is configured to use this AMD vLLM endpoint when the droplet is online; the provider pill at the top right reads live from /api/healthcheck/llm.
Optional product workflow
Have a real document? Try the full workflow
The hackathon path above uses seeded data so it always works. This upload surface shows where customer documents enter the same Triad Review pipeline.
Hosted demo uses sample documents. Private installs run on your own AMD MI300X (or any OpenAI-compatible provider).