Clex | Devpost

Inspiration and problem

A founder’s legal work starts before there is a contract to upload. Company formation, changing operations, hiring, and supplier relationships all depend on facts that generic checklists miss. Founders need a way to preserve those facts, see what remains uncertain, and prepare focused questions for qualified counsel.

What Clex does

Clex is a company-aware legal preparation workspace. A founder confirms their company’s stage, location, activities, and other context. Clex uses those answers to organize registration preparation and a starting checklist. As the company grows, separate employment, supplier, and other matter workspaces hold conversations and documents. In the supplier demo, Clex compares a sample conversation with a sample agreement, quotes the exact passages that differ, explains why the mismatch matters to that company, and helps prepare an editable response and a packet for a lawyer. A drafted response does not silently close an open finding; new evidence makes an earlier review stale.

How we built it

The prototype uses TypeScript, Next.js, React, Tailwind CSS, and PostgreSQL. Private documents and company data are scoped to a guest browser workspace with database row-level security. TXT, DOCX, and text-based PDF uploads feed bounded analysis. An instant local preparation review is available by default. A separate, explicit opt-in can send synthetic demo sources and company details to an OpenAI model; the response is schema-checked and quoted evidence must match source text. Failed model requests are shown as failures. We used Codex and Hoplite during development and can explain the code and workflow.

Tech and credits

TypeScript, Node.js, pnpm, Next.js, React, Tailwind CSS, PostgreSQL with row-level security, Zod, pg, Mammoth, pdf-parse, pdf-lib, Vitest, Playwright, ESLint, and Render. OpenAI Responses is used only for an explicitly consented model review; instant local preparation is the default. Codex and Hoplite were used as AI development tools.

Challenges and lessons

The hardest parts were keeping one company’s material isolated, grounding each finding in an exact excerpt, and keeping review state honest as sources change. Testing exposed a PDF parsing bundle issue and a live-model schema mismatch; both were corrected. The build reinforced that useful AI assistance needs visible evidence, recoverable errors, and a clear human review boundary.

Try it

Open the live preview and choose Click demo answers. The sample company, conversation, agreement, and certificate are synthetic. The full guided registration flow takes several minutes; the supplier matter is the shortest route to the evidence comparison. The preview is a hackathon prototype, not legal advice or a licensed lawyer. Jurisdiction packs are draft research pointers, calendar dates are planning placeholders, and generated drafts have not been lawyer-approved.