📐 MetroAI — KOLAS Compliance OS + Inverse Metrology Engine

Measurement uncertainty MCP server + 6 AI agents + verifiable audit trail + uncertainty-aware ML inverse metrology (11 instruments). ISO/IEC 17025 + KOLAS-accredited laboratories.

🇰🇷 한국어 요약 — MetroAI는 한국 KOLAS 인정기관(1,200+)의 시험·교정·표준물질 업무를 위한 AI 컴플라이언스 OS + 측정불확도 MCP 서버입니다. GUM/MCM 불확도 엔진, 6종 AI 에이전트, Ed25519+PROV-O 검증가능 감사추적, 역계측 엔진(11개 장비)을 제공하며, 자동화 테스트 243건과 정직성 규칙·Trustworthy AI 정책으로 AI 산출물의 신뢰성을 관리합니다. 상세 한국어 안내는 아래 섹션 참고.

What is this?

A compliance operating system for Korea's 1,200+ KOLAS-accredited testing, calibration, RMP, and inspection institutions — and a measurement uncertainty MCP server that any MCP-compatible AI client (Claude Desktop, Cursor, VS Code, etc.) can call directly.

As of v0.8.0 it also ships an inverse metrology engine (metroai.inverse): where the calibration templates compute uncertainty forward (inputs → U), the inverse engine recovers a parameter and its uncertainty from a measured signal (spectrum / image / diffraction) across 11 instruments, all sharing one GUM core and one ML-uncertainty core.

Built solo by Youngbum Kim (Ph.D.) — a practitioner from a KOLAS-accredited ISO 17034 reference-material producer who supported accreditation audits and got tired of redoing everything in Excel and email every quarter.

Related MCP server: timeseries-mcp

Quickstart — 30 seconds

As an MCP server (Claude Desktop, Cursor, etc.)

claude mcp add --transport http measurement-uncertainty \
  https://measurement-uncertainty.mcpize.run

→ then ask your AI: "Compute the GUM uncertainty for this voltage divider" or "Apply the TEM lattice template at 95% confidence".

As a web app (Streamlit)

pip install -e ".[dev,ml]"
streamlit run app.py

Live demo: metroai-gnbdv7pqq3quqsudb5pwvj.streamlit.app — hosted on Streamlit Community Cloud, which puts the app to sleep after inactivity; the first visit may ask for a one-click wake-up and take about a minute. A Hugging Face Spaces migration is planned (Roadmap P2). Or run locally with the two commands above.

As a Python library

pip install git+https://github.com/kyb8801/metroai.git

from metroai.templates import create_tem_lattice_calculator

calc = create_tem_lattice_calculator()
result = calc.calculate()
print(f"d = {result.measurand_value:.6f} nm "
      f"± {result.expanded_uncertainty:.4e} (k={result.coverage_factor:.2f})")

Inverse engine (NEW in v0.8.0)

from metroai.inverse import uncertainty, ml_inverse, INSTRUMENTS

# unified GUM budget — every instrument calls this
uc, rows = uncertainty.budget([("scan_calib", 1, 0.0016), ("noise", 1, 0.0001)])
print(uc, uncertainty.expand(uc, k=2))           # u_c, U(k=2)

# ML inverse + ML uncertainty (instruments that have a forward library)
mdl = ml_inverse.MLInverse().fit(X_train, y_train)
out = mdl.predict(X_query)                        # pred + epistemic std + conformal half-width

What MetroAI does

0. User-fit features (NEW in v0.7.0) — for KOLAS lab operators

After surveying the SEM-lab-operator journey end-to-end, four new features landed in v0.7.0 specifically to cover the applicant side of the accreditation workflow:

• 🔬 Domain-specific entry wizard — Landing page asks "Which instrument are you accrediting?" Five paths: SEM / TEM / AFM / OCD / general measurement. Each one routes to a domain dashboard with only the standards, KOLAS steps, uncertainty templates, and SOP checks that matter for that domain.

• 📚 Domain-specific KOLAS guides — Per-domain content: applicable ISO standards (4–6 per domain), six-step KOLAS accreditation process with typical pitfalls, 3–4 common nonconformities with root cause + MetroAI fix, and the typical uncertainty budget components. Content sourced from public ISO/SEMI/KOLAS-G-002 documents.

• 📝 KOLAS application form auto-generator — Fill an organization profile once → ReportLab generates a 7-section ISO/IEC 17025-style accreditation application PDF (organization info, scope, personnel, equipment, reference standards, environmental control, quality system). Generic template; final submission should be cross-checked against KAB's latest official form.

• 📋 Domain SOP rule-based checklist — Each domain ships with a 10-item SOP checklist derived from KOLAS-evaluator-perspective common findings. Real-time gap score and 1-click "add to orchestrator queue" for remediation.

End-to-end, the v0.7.0 changes raised our internal "lab-operator journey fit score" from 45% → 68% on a 7-stage scenario (entry → guide → form → KOLAS process → SOP check → simulation → end-to-end). The final 32% includes stages we can't automate ourselves (the "consulting + on-site evaluator hand-holding" piece of the journey).

1. Compliance OS — 6 AI agents (since v0.6.0)

Every agent output carries is_live / data_origin flags (live · stub · synthetic), so the UI can clearly distinguish authoritative data from heuristics.

2. Measurement uncertainty engine (since v0.5.0)

• GUM (ISO/IEC Guide 98-3) — symbolic partial derivatives, Welch–Satterthwaite, expanded U

• MCM (ISO/IEC Guide 98-3 Suppl. 1) — Monte Carlo with configurable n

• QMC — Sobol low-discrepancy sequence (verified ±0.003% agreement with GUM analytic on simple linear models)

• reverse_uncertainty — novel within prior-art search. Given a target combined U, compute the maximum allowed standard uncertainty per component. Not found in GUM Workbench, NIST Uncertainty Machine, or major open-source GUM tools as of 2026-05.

3. Nine calibration templates

4. Verifiable audit trail (NEW in v0.6.0)

• Ed25519 digital signatures (RFC 8032) — tamper-evident outputs

• W3C PROV-O provenance graphs (JSON-LD) — full input → model → output lineage

• Designed so a KOLAS auditor can verify no post-hoc tampering

5. Three MCP tools

6. Inverse metrology engine (NEW in v0.8.0) — metroai.inverse

The calibration templates (§3) run forward: given inputs, compute the uncertainty. The inverse engine runs the harder direction — given a measured signal (spectrum / image / diffraction), recover the parameter AND its uncertainty — across 11 instruments, all calling two shared cores so uncertainty and "AI" are consistent instead of ad-hoc per module.

Two shared cores

11 instrument modules (each calls the cores; ⟶ = ML hookup pending)

* The Lamb/acoustic module codes only public physics (e.g. Saviot & Murray 2009 breathing-mode relation); the inventive specifics live in a patent under KIPO review, not in this repo.

Honest scope (synthetic ≠ real; the figures are not inflated):

• ★★★ = real measured data. OCD on NIST L100P300; PL on MoS₂ A-exciton 1.850 ± 0.001 eV vs literature 1.85 eV. ★★ = synthetic or method-verified.

• OCD library inverse: error < 2 nm on NIST dies; naive and differential- evolution optimizers fail on the non-convex landscape (documented in flagship_v0_forward_inverse.py, not hidden).

• GPR reaches 0.19 nm noise-free but collapses to ~12.8 nm at 0.5 % noise (an overfit illusion, exposed in ocd_depth2p5_noise.py); KNN stays 3.5–3.8 nm across noise and is the robust choice.

• ML inverse needs a forward library (training data): strong for OCD/synthetic; PL and NSOM use peak-fit / clustering by design — the core picks the right tool per instrument rather than forcing a neural net everywhere.

• Inverse modules currently self-verify via __main__; pytest integration into the main CI suite is pending (tracked in the roadmap).

Honest metrics — kolas-audit-predictor

5-fold CV on synthetic data: accuracy 60.6% ± 3.1pp · ROC-AUC 0.628 ± 0.038 · Brier 0.241 · F1 0.636 (n=2000 × 6 features, label noise 0.15)

• GradientBoostingClassifier (n_estimators=200, depth=3, lr=0.05)

• Top 3 feature importances: months_since_last_audit (0.34) · personnel_turnover (0.25) · sop_completeness (0.24) — aligned with domain intuition.

• External validation on real KOLAS audit outcomes is pending. Synthetic-data metrics do not imply real-world accuracy.

• A prior sandbox figure of 87.1% has been removed from all artifacts. See docs/HONESTY_NOTES.md for citation rules.

Trustworthy AI

MetroAI treats AI-output trust as a first-class engineering problem — full policy in docs/TRUSTWORTHY_AI.md:

• Data-origin labels — every dataset and metric tagged real / synthetic / stub

• Honest metrics — publication rules in docs/HONESTY_NOTES.md (see synthetic-data caveats above)

• Verifiable audit trail — Ed25519 signatures + W3C PROV-O provenance graphs

• Uncertainty quantification — every inverse-engine estimate ships with its uncertainty

Standards compliance

• ISO/IEC 17025:2017 (testing & calibration laboratories)

• ISO/IEC Guide 98-3 (GUM) + Suppl. 1 (MCM)

• ISO 13528 + ISO 17043 (proficiency testing)

• ISO 18516 (microscope methods)

• ISO 25178-2 (areal surface texture)

• ISO 22489 (SEM-EDS quantitative)

• KOLAS-G-001 / G-002 (Korean accreditation guidelines)

• SEMI MF-1789 (OCD scatterometry)

• W3C PROV-O (audit provenance)

• PTB DCC 3.3.0 + D-SI 2.2.1 (digital calibration certificate — experimental dcc module)

• RFC 8032 (Ed25519 signatures)

DCC — Digital Calibration Certificate (experimental)

metroai.dcc is a first-pass (v1) toolkit for the PTB-led Digital Calibration Certificate XML standard (schema 3.3.0, wiki.dcc.ptb.de) in a KOLAS / ISO 17025 context.

What it does today — honest scope:

• dcc.parser — reads DCC XML: administrativeData (core data, items, lab, responsible persons, customer, statements) and measurementResults with D-SI values (si:real, si:realListXMLList, si:hybrid) and uncertainty in both the current measurementUncertaintyUnivariate form and the legacy expandedUnc form. Parsing is verified against an official PTB Good Practice example (bundled as a test fixture with its LGPL-3.0 notice retained).

• dcc.builder — turns a MetroAI GUMResult + certificate metadata (same cert_info keys as the KOLAS PDF exporter) into a schema-3.3.0 DCC draft: all schema-required elements, expanded uncertainty as the non-deprecated expandedMU, and an optional per-component uncertainty-budget table. Builder output validates against the official 3.3.0 XSD (checked with xmlschema in an optional test). Missing inputs become placeholders and are recorded in DCCBuilder.warnings — a draft with warnings is not a certificate.

• dcc.units — common unit strings → D-SI notation ("mm" → \milli\metre, "°C" → \degreecelsius); unmapped units are kept verbatim with a warning.

• metroai/dcc/kolas_map.md — mapping table: KOLAS / ISO 17025 §7.8 certificate content requirements ↔ DCC 3.3.0 elements, including known gaps.

from metroai.dcc import export_dcc_xml, parse_dcc

xml = export_dcc_xml(result, {"cert_number": "KOLAS-2026-...", "cal_org": "..."})
doc = parse_dcc(xml)          # round-trips: values, units, U, k, p survive

Not implemented yet (roadmap): XML-DSig signatures (a DCC without a signature is only a draft), alignment with the PTB basic_*/gp_* refType vocabulary, multi-item / before-after adjustment results, attachments (byteData), DCC 3.4 (release candidate as of 2026-07). Optional XSD validation needs pip install -e ".[dcc]" (installs xmlschema).

Streamlit app — v2-spec pages

v2 backbone (since v0.6.0):

1. 🏠 Landing (app.py) — KOLAS Compliance OS positioning + domain wizard

2. 🤖 6 Agents Dashboard (pages/11) — Quality Manager daily view, KPI strip + task queue

3. 📋 SOP Gap Analyzer (pages/12) — Technical Manager work surface, AI-detected gaps + v0.7 domain-specific checklist

4. 📰 KOLAS Feed (pages/13) — kolas-monitor regulatory news

5. 🎯 Audit Risk Detail (pages/14) — explainability, waterfall + AI reasoning + what-if

6. 📅 Ops Backbone (pages/15) — certificates / personnel / schedule

v0.7.0 P0 — lab-operator journey (NEW):

7. 🔬 SEM domain dashboard (pages/16) — SEM-EDS standards + KOLAS process + nonconformities + SOP checklist

8. ⚛️ TEM domain dashboard (pages/17) — lattice constant, ISO 29301 + Cs-corrector spec

9. 📐 AFM domain dashboard (pages/18) — surface roughness Sa/Sq/Sz per ISO 25178-2

10. 📏 OCD domain dashboard (pages/19) — Scatterometry / RCWA library matching per SEMI MF-1789

11. 📝 KOLAS application form (pages/20) — Fill-once → 7-section ISO 17025-style PDF (KAB-F-21 reference)

Plus the legacy v0.5 calibration / PT / certificate pages (pages/1–10).

Repository layout

metroai/
├── app.py                     ← v2-spec landing page (Streamlit entry)
├── pages/                     ← Streamlit multi-page
│   ├── 1_📐_불확도_계산.py     ← Uncertainty calculator (KR)
│   ├── 2_📊_PT_분석.py         ← PT analysis (KR)
│   ├── 3_📄_교정성적서.py      ← Calibration certificate PDF
│   ├── 4_🔄_불확도_역설계.py    ← Reverse uncertainty (novel)
│   ├── 11_🤖_6_Agents.py       ← v2 block 2: main dashboard
│   ├── 12_📋_SOP_갭_분석.py     ← v2 block 4: SOP gap analyzer
│   ├── 13_📰_KOLAS_피드.py      ← v2 block 5: regulatory feed
│   ├── 14_🎯_감사_위험_상세.py   ← v2 block 3: risk explainability
│   └── 15_📅_인증서_인력_일정.py ← v2 block 6: operations
├── metroai/
│   ├── core/                  ← GUM / MCM / model parsing
│   ├── agents/                ← 6 AI agents backbone
│   ├── audit/                 ← Ed25519 + PROV-O
│   ├── connectors/            ← KOLAS / DART / NTIS live fetch + stub fallback
│   ├── math/                  ← Sobol QMC
│   ├── ml/                    ← GBT audit-risk model + synthetic data
│   ├── templates/             ← 9 calibration templates
│   ├── dcc/                   ← NEW v0.8.x: DCC (digital calibration
│   │                              certificate) parser/builder — experimental
│   ├── inverse/               ← NEW v0.8.0: uncertainty-aware ML inverse metrology
│   │   ├── __init__.py        ←   package: cores + INSTRUMENTS map (11)
│   │   ├── uncertainty.py     ←   ① unified GUM core
│   │   ├── ml_inverse.py      ←   ②③ ML inverse + ML uncertainty core
│   │   ├── metrology_module_2..10_*.py ← 9 instrument modules
│   │   │                           (XRR/TEM lattice/Raman/TEM strain/SEM/AFM/PL/NSOM/Lamb)
│   │   ├── flagship_v0_forward_inverse.py ← OCD forward+library inverse (R+T=1, err<2nm)
│   │   ├── flagship_v1_autodiff_gpu.py    ← OCD autodiff inverse (Meent torch)
│   │   ├── ocd_depth1..2p6_*.py  ←   OCD accuracy / GPR / noise-robustness deep-dive
│   │   ├── nist_real_data_inverse.py ← NIST L100P300 real-die inverse
│   │   └── PLATFORM_INDEX.md   ←   inverse engine map + honest status
│   ├── schemas.py             ← Pydantic v2 input validation
│   ├── exceptions.py          ← MetroAIError hierarchy
│   └── mcp_server.py          ← MCP stdio server
├── tests/                     ← 243 unit tests (pytest)
├── docs/
│   ├── HONESTY_NOTES.md       ← Citation rules
│   ├── TRUSTWORTHY_AI.md      ← Trust & governance policy
│   ├── v0.7.0_ROADMAP.md      ← Next 3 months
│   └── RELEASE_NOTES_v0.6.0.md
├── mcp_manifest.json          ← MCPize manifest (v0.6.0)
└── pyproject.toml

Roadmap (v0.7.0 → v0.8.0 — 2026-05 → 2026-08)

Reordered 5/19 around the lab-operator journey (after a virtual-user audit revealed v0.6.0 covered only 45% of the path-to-accreditation). Philosophy shifted from outbound-first → user-fit-first.

See docs/v0.7.0_ROADMAP.md for the full plan.

Tech stack

• Python 3.10+ (tested on 3.10 / 3.11 / 3.12)

• Streamlit (web UI)

• sympy / numpy / scipy (numerical)

• Pydantic v2 (input validation)

• cryptography (Ed25519)

• scikit-learn (GBT model + inverse ML cores, optional [ml] extra)

• refnx / meent (XRR / OCD inverse, optional)

• reportlab + openpyxl (PDF + Excel export)

• altair / plotly (visualizations)

Tests

pip install -e ".[dev,ml]"
pytest tests/ -v

Latest CI on Python 3.10 / 3.11 / 3.12 — 243 passing tests across the v0.5 → v0.8 suites (+1 optional DCC XSD-validation test that runs only when xmlschema and a local copy of the official schema are present). Inverse modules (metroai/inverse/) currently self-verify via their __main__ blocks; folding them into the pytest CI suite is a P1 roadmap item.

License

MIT License. See LICENSE.

Community

• GitHub Issues — bug reports, feature requests

• GitHub Discussions — Q&A, design discussion

• MCPize page — install + reviews: mcpize.com/mcp/measurement-uncertainty

• Glama listing — glama.ai/mcp/servers?query=metroai

• Email — kyb8801@gmail.com (KOLAS-side feedback especially welcome)

한국어 사용자를 위한 안내

본 프로젝트는 한국 KOLAS 인정 기관 실무자가 직접 사용할 수 있도록 한국어 페이지와 한국어 UI 를 지원합니다. 자세한 한국어 가이드는 docs/RELEASE_NOTES_v0.6.0.md 및 Streamlit 앱의 한국어 페이지들 (불확도 계산 / PT 분석 / 교정성적서 / 불확도 역설계 / KOLAS 로드맵 / 6 Agents 대시보드 / SOP 갭 분석 / KOLAS 피드 / 감사 위험 상세 / 운영 백본) 을 참고해주세요. cold-feedback 환영합니다 — kyb8801@gmail.com.

v0.8.0 부터는 역계측 엔진(metroai.inverse)이 추가되었습니다 — 측정 신호(스펙트럼· 이미지·회절)로부터 파라미터와 그 불확도를 동시에 복원하며, 11개 장비가 공용 GUM 코어와 ML-불확도 코어를 함께 호출합니다. 합성·실측 등급(★)과 한계(GPR 노이즈 붕괴, ML 적용 범위)를 README 에 정직하게 표기했습니다.

Built with care by @kyb8801 · KOLAS RMP operations background.