Korean Public Procurement Analytics

Overview

A knowledge product and workflow platform built for Korean public procurement practitioners preparing for the 공공조달관리사 (Public Procurement Manager) national credential. The free layer is a static reference site (Astro Starlight) with 82 concept cards covering the 14 core procurement regulations — each card law-verified via live law.go.kr API cross-check. The paid layer is a FastAPI practitioner engine that queries 나라장터 (G2B) and 지방재정365 (LOFIN) APIs, applies statutory compliance rules (낙찰하한율, 적격심사 scoring, 수의계약 thresholds), and identifies anomaly patterns in 24,338 local government contracts — surfacing procurement decisions that deviate from legal guardrails.

Problem

Two problems run in parallel. For the credential market: the 공공조달관리사 exam expects practitioners to apply 14 statutory instruments to realistic tender scenarios. No organized, searchable, law-linked reference exists. Exam prep relies on paper study guides that go stale as regulations change annually. For the analytics market: Korean local government procurement data is published through 지방재정365 (LOFIN) as monthly parquet files with no standard anomaly detection layer. A city comptroller or procurement auditor who wants to know whether last quarter's 수의계약 awards follow the legal threshold criteria has no off-the-shelf tool. Manual review across 24,338 contracts is not viable.

Constraints

• G2B API (나라장터) and LOFIN API rate limits require aggressive caching — brute-force re-fetching against 나라장터's 40M+ contract records is not viable on free-tier credentials
• Statutory rules change annually (행정안전부 고시): 낙찰하한율 rates and 표준노무비 rates embedded in the compliance engine must be updated at the start of each fiscal year
• 82 public concept cards pass the public-sync quality gate (title, law_id, verified_at, jurisdiction fields required); remaining 22 are private or unresolved — MISMATCH status from law.go.kr verification blocks public sync
• 지방재정365 data covers local government (지자체) contracts only; 나라장터 central government data is separately gated and requires higher API tier for bulk download
• Outreach to 지자체 procurement offices requires named contacts before any email — cold emails without a named recipient are blocked by hard rule

Approach

Free layer first: extract 99 concept cards from a 15-chapter exam-prep book via Anthropic Files API (claude-sonnet-4-6 as OCR + semantic extractor), populate law_id fields against MST code registry (28 instruments Tier 1–3), run live law.go.kr verification for each article citation, run public-sync quality gate, deploy to Astro Starlight on Vercel. Paid layer: DuckDB/Parquet pipeline querying G2B and LOFIN APIs; compliance functions compute 낙찰하한율 floor, 적격심사 score bracket, and 수의계약 threshold per contract record; anomaly detection applies isolation forest and local outlier factor for price deviations and contract-splitting patterns. Output: FastAPI + Jinja2 demo with 5 client profiles deployed to Railway.

Key Decisions

Tech Stack

• Python ≥3.11, uv
• FastAPI + Jinja2 (practitioner web app)
• Astro Starlight (free layer static site)
• DuckDB, pandas, PyArrow, Parquet
• 나라장터 G2B API (공공데이터포털 G2B_API_KEY)
• 지방재정365 LOFIN API (LOFIN_API_KEY)
• law.go.kr Open API (LAW_OC) — concept card verification
• Anthropic SDK (claude-sonnet-4-6) — book OCR ingest
• scikit-learn (isolation forest, LOF, gradient boosting price model)
• Railway (demo deployment), Vercel (Astro static site)
• pytest (618 tests)

Result & Impact

• 618
  Tests passing
• 24,338
  Local government contracts analyzed
• 6
  Live data runs
• 82
  Public concept cards (law-verified)
• 28 (14 core + 14 supporting)
  Statutory instruments indexed
• 15 (full exam-prep book)
  OCR chapters ingested

The first open-access, law-linked Korean procurement reference with automated regulatory drift detection. Each concept card is verified against live law.go.kr article text — not a static PDF. The paid layer applies statutory compliance rules (낙찰하한율, 적격심사, 수의계약 thresholds) to actual 지자체 contract data, surfacing deviations a human auditor would need a week to identify manually.

Learnings

• Law-linking is the moat, not content. Procurement exam prep content is commoditized (PDF study guides exist). Automatically verifiable, article-level citations that survive regulatory updates are not. The verify_card_accuracy.py MATCH/MISMATCH feedback loop makes the product defensible in a way a static guide cannot be.
• Statutory rule encoding is brittle at the edges. The 낙찰하한율 rate table has seven category × estimate × issuer combinations, each with sub-percentage precision. Encoding these correctly required reading the 행정안전부 고시 original text, not a secondary source. One digit error in a rate table invalidates an entire compliance bracket.
• Demo realism determines outreach conversion. The Railway demo with 5 real BRNs and live 나라장터 API field names (MAS 2단계경쟁, 영업일 countdown, 복수예비가격 15-dot spread) converted a practitioner meeting from a cold email. A generic 'analytics dashboard' would not have.
• Named-contact outreach is a hard constraint, not a soft preference. Every 지자체 outreach campaign stalls at the same point: emails without a named recipient go to general inboxes and die. The pipeline is call-first (get a named contact) → PDF + data package → email. Reversing this order wastes effort.