I am an AI student. I build AI agent systems that connect model reasoning to operational workflows through evaluation, human approval, and auditable execution.
My focus is the layer between a model's output and a real system action: what the agent is allowed to do, who approves it, and what record is left behind.
Problem — When an LLM agent operates an ERP system, model reasoning alone decides when inventory and purchasing records change. Functional authorisation (a user may access a feature) is not the same as authorisation for an agent to execute that action autonomously.
System approach — External supply-chain risk analysis feeds a three-tier workflow: L1 observation (read-only risk KPIs, heatmap, alerts), L2 proposal (impact analysis, what-if comparison, immutable proposals submitted for review), and L3 approval and execution (review of approval evidence, approve or reject, gateway-mediated execution, audit timeline). Protected ERP writes require human approval before execution. Proposal, Approval, and Execution are stored as separate records alongside audit entries and idempotent execution receipts.
Verifiable evidence — 327 automated tests passing on the public default
branch; tagged v0.1 and v1.0 GitHub Releases with English release notes;
a published procurement-flow diagram with its editable draw.io source; a
documented known-limitations section.
Scope and limitations — Proof of concept, not a deployed system. One SQLite database represents a single organisation, so this is not row-level multi-tenancy on a shared database. The application-layer audit trail is tamper-evident but does not defend against someone with host or database administrator access. SQLite transaction guarantees do not extrapolate to external ERP APIs; cross-system execution would still require an outbox, worker, and reconciliation strategy. Demo accounts and synthetic data are for local evaluation only.
🔗 falltwo/AI-Risk-Based-Inventory-ERP
Problem — A first-pass contract review means reading clauses one by one and manually cross-checking them against legal sources, which is slow and hard to trace back to a citation.
System approach — Retrieval-augmented generation over uploaded legal documents, combined with a multi-tool agent workflow that routes between document retrieval, knowledge-base question answering, and external legal-source lookup. Answers are grounded in retrieved chunks and carry citations. The repository includes evaluation datasets and batch validation routes so output quality can be measured rather than only demonstrated, plus a Streamlit demo path and a FastAPI + Vue service path.
Verifiable evidence — MIT licence; CI workflow defined in the repository; evaluation API routes and evaluation datasets present in the public tree; public repository topics covering RAG, agents, and retrieval.
Scope and limitations — This is a retrieval and drafting aid, not legal advice. It does not establish legal accuracy or regulatory compliance, and its output requires review by a qualified person before use. I have no measured business outcomes to report for it.
🔗 falltwo/Contract-compliance-agent
These are directions I am reading about and experimenting with. They are not claims about what the projects above already implement.
- Governance controls that hold when the underlying model is not trusted
- Evaluation methods for agent systems beyond single-response benchmarks
- Long-running observability and cost control across multi-model routing
- Human-in-the-loop approval design: what to surface to a reviewer, and when
+ LiteLLM for the model layer · Streamlit · FastAPI · SQLite
- Email — gjui890@outlook.com
- Location — Taiwan
