01
Summary
One verdict for the procedure, with the blockers ranked by severity underneath and a readiness score labelled as indicative, not as a promise.
The problem
The expensive mistakes are never on the first page
A Romanian construction firm deciding whether to bid reads several hundred pages under a two week deadline: eligibility criteria, evidence requirements, penalty clauses, payment schedules, guarantee obligations, technical specifications. The work falls to the most senior people in the company, it is done by hand, and the failure modes are ordinary rather than exotic. A qualification document nobody noticed was required. A payment term that quietly makes the project cash flow negative from month one. A penalty clause that turns a two week delay into a loss.
82% of Romanian public contracts are awarded on lowest price, against a European average of 54%. Every incentive pushes the offer down; the risk sits in signing something the company cannot fund. Corthea exists to make that risk visible before submission instead of after.
What it does
Six workspaces, one thread from the documentation to the decision
The platform is not a chat window over a pile of PDFs. It is a sequence a bidding team can actually follow, where each step produces something the next step consumes, and every conclusion stays attached to the paragraph that produced it.
02
Requirements
Every requirement in the documentation, linked to the evidence the company can produce and to the document being prepared for the bid file.
03
Contract money
Working capital, guarantees, the month the balance turns negative, penalties, financing cost. Deterministic engines, not generated numbers.
04
Compliance
Coverage, gaps and open questions, with detected status and human confirmed status stored separately.
05
File and DUAE
Submission checklist and a European Single Procurement Document skeleton, parsed from the authority's own request file.
06
Clarifications
Draft questions to the contracting authority, each one carrying the source paragraph that made it necessary.
The screens
From the running system, unretouched
The interface is in Romanian because the users are Romanian contractors and the source documents are Romanian public procurement files. What matters in each screenshot is not the layout, it is the discipline visible in the numbers.
Under the hood
The half a client never sees, which is the half that decides whether they can use it
Anyone can put a language model behind an upload button. What takes seven months is everything that has to be true before a company will let that thing near a procurement file: isolation that holds, money that cannot run away, an answer to the question of who did what, and a way back when something breaks.
Tenant isolation lib/db
Isolation is enforced by PostgreSQL, not by remembering to add a filter in application code. Every request runs inside a tenant context; a query that forgets its tenant returns nothing rather than someone else's documents.
- 59 tables, 52 of them carrying row-level security policies
- Tenant context set per connection, not per query
- Seat management and invitations scoped to the tenant
- A break-glass path for support access, which writes an audit record when used
Entitlements three layers, server side
What a plan allows is decided on the server every time, in three cooperating layers, so a limit cannot be lifted by editing anything the browser can reach.
entitlementsfor the product surface a tenant has boughtrotix-server-entitlementsfor platform level ceilingsofferiq-server-entitlementsfor module specific gates- Trial gates and plan boundaries evaluated in the same path as production access
Document pipeline lib/documents, lib/indexer
Ingestion handles what real companies actually send: scans, legacy Office formats, archives, mixed quality PDFs. Everything is validated before it is opened.
- MIME validation and path safety on every upload, no trusting file names
- OCR for scanned pages, with scan quality detected rather than assumed
- Legacy Office and archive expansion, so a ZIP of DOCs is not a dead end
- Indexing into a knowledge layer with tiers, pointers and citation cards
- PDF font and embedded image handling for documents that resist extraction
Deterministic engines lib/contract-money
Every number that touches money is computed in ordinary code, reproducibly, and can be recalculated by hand. No financial figure in this product is generated by a language model.
- Monthly cash flow with advance, retention and payment term as inputs
- Bid bond and performance guarantee, with locked capital tracked separately
- A penalty and interest engine driven by the clauses read from the contract
- The financial stress point: the month the project first requires outside funding
- Assumptions marked as assumptions, everywhere, in the data model itself
Cost governance lib/cost-caps
Model spend is metered per tenant and checked before the call is made, not reconciled afterwards. At the ceiling the system refuses work and says so.
- Per tenant metering with a monthly ceiling, fail closed by design
- Usage accounting kept separately from billing state
- Rate limiting in two layers, one static and one runtime aware
- Search fails closed when authorisation is uncertain, covered by its own test
Security lib/auth, lib/totp
Access control was built at the start rather than added before launch, which is the only time it is cheap.
- Two factor authentication with TOTP, and sessions that can be revoked individually
- Secrets encrypted at rest with a dedicated crypto path
- Invite only registration, with email verification and password reset flows
- An append-only audit trail across authentication, documents and money
Compliance lib/account-deletion
The obligations a European customer will ask about in the first call, implemented rather than promised.
- GDPR data export on demand, per account
- Account deletion that actually removes the tenant's data
- Public share links that are scoped and revocable
- Sanitisation on anything leaving the tenant boundary
Operations lib/migrations, backups
The part that decides whether a system survives its first bad day.
- Backup and restore exercised for real on 19 July 2026, not assumed
- Versioned migrations across 17 SQL files, applied in order
- A recovery dossier: written procedure for getting the system back
- Resilience paths so one failing subsystem does not take the analysis down
Specification
Counted, not estimated
| Measure | Value | How it was counted |
|---|---|---|
| Application code | 118,590 lines | TypeScript and TSX in app/ and lib/ |
| Source files | 473 | Same two directories |
| API endpoints | 187 | Route handlers under app/api/ |
| Pages and components | 98 | TSX files outside the API layer |
| Database tables | 59 | Schema, with 52 carrying row-level security policies |
| Schema migrations | 17 files | Versioned SQL, applied in order |
| Commits | 242 | February to September 2026 |
| Build time | 7 months | One person, part time, with AI assistance |
What I do not claim
The part most portfolios leave out
This section exists because it is the part a serious buyer actually needs, and because a vendor who cannot tell you where their system is weak has not looked.
- Corthea has no paying customers. It is a production grade system without a commercial launch behind it.
- Its extraction accuracy has not been benchmarked on a scored set of real Romanian tender files. The infrastructure is finished; that measurement is not.
- No figure exists for money or time saved, because none was measured. You will not find a percentage on this page.
Which is exactly why every project I deliver includes a scored evaluation on the client's own documents, before launch, instead of a promise of accuracy afterwards.
What it means for you
Corthea is not for sale. It is the reference for what gets built into your product.
If you are a software company, an agency or a product team with a document problem, the relevant fact is not that this platform exists. It is that one person took an unstructured, high stakes document domain and delivered a working system with real access control, cost governance and an audit trail, then wrote down honestly what it does not prove.
The reusable half of it, ingestion, indexing, page level citations, the human review screen, per tenant metering, is the library every client project starts from. That is why a document pipeline from me takes weeks rather than months, and why the boring parts are already solved when we start.
Figures are counted from the repository and the database schema on 10 September 2026, not estimated. Screenshots come from the running system and are unretouched. The 82% and 54% procurement figures are European Commission Single Market Scoreboard data for 2024.
Rotix AI Systems. Document intelligence, web applications and internal tools, built in TypeScript and delivered with tests, documentation and a handover call.