Latvian tender matcher: the LLM structures text, deterministic code decides compliance

It parses free-text Latvian public-tender requirements into structured constraints and checks them against a product catalogue with deterministic rules. It gives a per-criterion verdict, has a FastAPI review UI and exports an Excel report. 704 tests pass; in production since 2026-09-12.

Problem

Tender rows state requirements in free Latvian prose, with decimal commas, operator phrases (ne mazāk kā, no X līdz Y) and inconsistent units. Answering a tender by hand means re-reading the catalogue for every row, and a wrong "compliant" is a safety and bid-integrity risk.

Constraints

• False positives are treated as strictly worse than false negatives.
• The LLM never decides compliance. Numbers, units and operators are handled by code, never by a model.
• An explicitly stated range is a hard bound with zero tolerance. Device-fit attributes are exact. Tube length is +/-10 cm absolute.
• The catalogue was transcribed by hand from a supplier PDF: 33 families / 43 nutrition variants, 21 families / 136 device variants.
• Production runs Python 3.12.3. The docs claimed 3.13, so the floor was corrected and a test enforces it.

What I built

 tender.xlsx --> Excel loader --+
                                 +--> extraction        requirement parser, device parser,
 products.yaml -> YAML loader --+     (LLM: text -> structure only)   item-numbering hierarchy
                                         |
                                         v
                                 normalisation       units, decimal comma,
                                                     "ne mazāk kā" -> >=, "līdz" -> <=, range
                                         |           controlled vocabulary
                                         v
                                 matching engine     deterministic, per criterion:
                                 tolerance policy.yaml  COMPLIANT | PARTIALLY_COMPLIANT
                                         |              NON_COMPLIANT | UNKNOWN
                        +----------------+----------------+
                        v                                 v
                 Excel report               FastAPI + Jinja2 review UI
                                            upload, results, per-item detail,
                                            product browser

 library: uploaded tenders keyed by SHA-256 of the file bytes
 verdicts are never stored; they are recomputed on every view

• Set the design rule that the LLM structures text and never answers "does this match?", and wrote it into the project rules.
• Built the Latvian operator/unit normaliser, the policy-driven tolerance table and the deterministic engine.
• Chose the next work by measured return rather than by the backlog, and killed 7 items on evidence.
• Deployed to production and matched the supported Python floor to production, with a test enforcing it.

Outcome

Stack

Redacted screenshot