ArchiveYVEA case study (2022-2025), kept in its original layout.
Untangling regulatory complexity with AI
A strategic retrospective: how I turned an administrative nightmare into an automated audit engine for international trade.
Author’s note
International trade is the last stronghold of administrative opacity. With YVEA, my goal was not to build a technological gadget but to lift a structural constraint: the time it takes to reach qualified information. This is the story of that venture, from field immersion to the final strategic call.
From chaos to opportunity
Exporting is a race against the clock. Embedding myself with export administration managers, I identified three pillars of a systemic administrative nightmare:
The product insight
The real bottleneck was not regulatory, it was temporal. Every hour of uncertainty is a straight loss on the product’s gross margin.
The pivot: from interface to AI
We first built a management interface (SPoT). But flow analysis revealed that the breaking point was manual semantic comparison.
Decision: turn YVEA into an AI orchestration engine.
A robust SaaS architecture
The state machine
To escape the straitjacket of frozen PDF forms, I architected the product around rigorous state management. That is what makes error recovery and asynchronous work possible.
- Draft
- Submitted
- In review (AI)
- Approved
Auditability
Every file carries its own contextual thread, tied to its certificate identifier.
Real time
Instant notifications, to cut the dead time between an error and its fix.
Zero download
High-fidelity preview, so a file can be inspected without leaving the interface.
It is night and day. I used to block out a whole morning to check a complex file. Today YVEA does 90 % of the work for me. I only rule on the disputed cases. I went from two hours to five minutes per file.
AI orchestration
The long-document problem
A PVoC report can run past 40 pages. To work around the token limit, I supervised a semantic chunking strategy with overlap.
01. Hybrid pipeline
To hold a 92 % success rate, we combine native extraction, fast and free, with resilient OCR, slow and costly, for degraded scans.
02. Governance and cost
Model segmentation: small models classify, large ones reason. Cost held between $0.15 and $0.30 per file.
Technical stack
- Next.js 14
- TypeScript
- Tailwind CSS
- Zustand
- Python (FastAPI)
- LangChain
- OpenAI GPT-4o
- Azure Doc Intel.
- PostgreSQL
- Supabase
- Vercel Edge
- Redis
- GitHub CI/CD
- Jira / Linear
- Unit testing
What I would do differently
01. AI is no longer the moat
Reasoning has become a commodity. Given another run at it, I would bet less on the power of algorithmic orchestration and more on capturing exclusive proprietary data.
02. Vertical SaaS
The real competitive moat lies in an integration so deep into the user’s daily work that a general-purpose model cannot dislodge it.
Facing something similar?
Let us talk on LinkedIn