MetroAI β KOLAS Compliance OS
by kyb8801
README.md
# π MetroAI β 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.
[](https://github.com/kyb8801/metroai/actions)
[](LICENSE)
[](https://www.python.org)
[](https://github.com/kyb8801/metroai/releases)
[](https://github.com/kyb8801/metroai/actions)
[-FF4B4B)](https://metroai-gnbdv7pqq3quqsudb5pwvj.streamlit.app)
[](https://github.com/punkpeye/awesome-mcp-servers/pull/6980)
[](https://mcpize.com/mcp/measurement-uncertainty)
[](https://glama.ai/mcp/servers/kyb8801/metroai)
---
> **π°π· νκ΅μ΄ μμ½** β MetroAIλ νκ΅ KOLAS μΈμ κΈ°κ΄(1,200+)μ μνΒ·κ΅μ Β·νμ€λ¬Όμ§ μ
무λ₯Ό μν
> **AI μ»΄νλΌμ΄μΈμ€ OS + μΈ‘μ λΆνλ MCP μλ²**μ
λλ€. GUM/MCM λΆνλ μμ§, 6μ’
AI μμ΄μ νΈ,
> Ed25519+PROV-O κ²μ¦κ°λ₯ κ°μ¬μΆμ , μκ³μΈ‘ μμ§(11κ° μ₯λΉ)μ μ 곡νλ©°, μλν ν
μ€νΈ 243건과
> [μ μ§μ± κ·μΉ](docs/HONESTY_NOTES.md)Β·[Trustworthy AI μ μ±
](docs/TRUSTWORTHY_AI.md)μΌλ‘ AI μ°μΆλ¬Όμ μ λ’°μ±μ κ΄λ¦¬ν©λλ€.
> μμΈ νκ΅μ΄ μλ΄λ [μλ μΉμ
](#νκ΅μ΄-μ¬μ©μλ₯Ό-μν-μλ΄) μ°Έκ³ .
## What is this?
A **compliance operating system** for Korea's 1,200+ KOLAS-accredited testing,
calibration, RMP, and inspection institutions β 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 β 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.) β 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.
---
## Quickstart β 30 seconds
### As an MCP server (Claude Desktop, Cursor, etc.)
```bash
claude mcp add --transport http measurement-uncertainty \
https://measurement-uncertainty.mcpize.run
```
β 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)
```bash
pip install -e ".[dev,ml]"
streamlit run app.py
```
Live demo: [metroai-gnbdv7pqq3quqsudb5pwvj.streamlit.app](https://metroai-gnbdv7pqq3quqsudb5pwvj.streamlit.app) β 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
```bash
pip install git+https://github.com/kyb8801/metroai.git
```
```python
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"Β± {result.expanded_uncertainty:.4e} (k={result.coverage_factor:.2f})")
```
### Inverse engine (NEW in v0.8.0)
```python
from metroai.inverse import uncertainty, ml_inverse, INSTRUMENTS
# unified GUM budget β 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) β 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:
- **π¬ Domain-specific entry wizard** β 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.
- **π Domain-specific KOLAS guides** β Per-domain content: applicable ISO
standards (4β6 per domain), six-step KOLAS accreditation process with
typical pitfalls, 3β4 common nonconformities with root cause + MetroAI fix,
and the typical uncertainty budget components. Content sourced from public
ISO/SEMI/KOLAS-G-002 documents.
- **π KOLAS application form auto-generator** β Fill an organization profile
once β 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.
- **π Domain SOP rule-based checklist** β 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% β 68%** on a 7-stage scenario (entry β guide β form β
KOLAS process β SOP check β simulation β 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 β 6 AI agents (since v0.6.0)
| Agent | Role | Data source |
|---|---|---|
| `semi-intel` | Semiconductor industry signals | DART (Korea FSS) + NTIS R&D feeds |
| `job-scout` | Personnel turnover signal | Public job postings (baseline stub) |
| `kolas-monitor` | KOLAS / KAB / KTR notice scan | knab.go.kr live fetch (with stub fallback) |
| `kolas-audit-predictor` | Next-audit risk prediction | Rule baseline + optional GBT model |
| `orchestrator` | Integrated P0/P1/P2 task queue | All other agents |
| `schedule` | Calibration / audit / review calendar | Internal events DB |
**Every agent output carries `is_live` / `data_origin` flags** (live Β· stub Β·
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) β symbolic partial derivatives, WelchβSatterthwaite, expanded U
- **MCM** (ISO/IEC Guide 98-3 Suppl. 1) β Monte Carlo with configurable n
- **QMC** β Sobol low-discrepancy sequence (verified Β±0.003% agreement with GUM analytic on simple linear models)
- **`reverse_uncertainty`** β *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
| Template | Domain | Standard |
|---|---|---|
| `gauge_block` | Length | KOLAS-G-002 |
| `mass` | Mass (weights) | OIML R 111 |
| `temperature` | Temperature (PRT) | ITS-90 |
| `pressure` | Pressure | KOLAS-G-002 |
| `dc_voltage` | DC voltage | KOLAS-G-002 |
| `tem_lattice` (v0.6 new) | TEM d-spacing | Si CRM reference |
| `sem_eds` (v0.6 new) | SEM-EDS quantitative | ZAF, ISO 22489 |
| `afm_roughness` (v0.6 new) | AFM surface roughness Sa/Sq | ISO 25178-2 |
| `ocd_scatterometry` (v0.6 new) | OCD CD measurement | RCWA, SEMI MF-1789 |
### 4. Verifiable audit trail (NEW in v0.6.0)
- **Ed25519** digital signatures (RFC 8032) β tamper-evident outputs
- **W3C PROV-O** provenance graphs (JSON-LD) β full input β model β output lineage
- Designed so a KOLAS auditor can verify no post-hoc tampering
### 5. Three MCP tools
| Tool | Use case |
|---|---|
| `calculate_uncertainty` | GUM calculation across the 9 templates |
| `pt_analysis` | Proficiency Testing β z-score / En / zeta per ISO 13528 + 17043 |
| `reverse_uncertainty` | Target-U β per-component limit allocation |
### 6. Inverse metrology engine (NEW in v0.8.0) β `metroai.inverse`
The calibration templates (Β§3) run *forward*: given inputs, compute the
uncertainty. The **inverse engine** runs the harder direction β **given a
measured signal (spectrum / image / diffraction), recover the parameter AND
its uncertainty** β across 11 instruments, all calling **two shared cores** so
uncertainty and "AI" are consistent instead of ad-hoc per module.
**Two shared cores**
| Core | File | Role | Verified (sandbox) |
|---|---|---|---|
| β Unified GUM | `inverse/uncertainty.py` | `combine_gum` Β· `expand` Β· `monte_carlo` Β· `budget` Β· `sensitivity_fd` | block-gauge u_c = 0.0594 mm, U(k=2) = 0.1189; MC cross-check 0.0595 (match) |
| β‘β’ ML inverse + ML uncertainty | `inverse/ml_inverse.py` | RandomForest ensemble (epistemic std) + conformal (distribution-free) + `combine_with_gum` | conformal 90% target β 89% empirical coverage |
**11 instrument modules** (each calls the cores; βΆ = ML hookup pending)
| Instrument | Method | Uncertainty | AI | Grade | Data |
|---|---|:---:|:---:|:---:|---|
| OCD scatterometry | RCWA (Meent) + library | β
GUM | β
KNN/GPR + conformal | β
β
β
| NIST L100P300 (real) |
| PL / exciton | peak fit | β
curve_fit | peak fit | β
β
β
| PhD Valley data (real) |
| XRR | Parratt/AbelΓ¨s (refnx) | β
covariance | βΆ | β
β
| synthetic |
| TEM lattice | windowed FFT + subpixel | β
GUM budget | βΆ | β
β
| HRTEM |
| TEM strain | Geometric Phase Analysis | β
| βΆ | β
β
| GPA |
| SEM CD | threshold + PSF | β
| βΆ | β
β
| synthetic |
| AFM roughness | ISO 25178 Sa/Sq/Sz | β
GUM budget | βΆ | β
β
| real .spm |
| NSOM | hyperspectral + k-means | β
GUM budget | β
k-means | β
β
| PhD ipynb |
| Lamb acoustic | breathing-mode fβ + 4D | β
GUM budget | β
Mahalanobis | β
β
| public physics only* |
| Raman | Lorentzian quant | β
curve_fit | βΆ | β
β
| synthetic |
\* 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 β real; the figures are not inflated):
- β
β
β
= real measured data. OCD on NIST L100P300; PL on MoSβ A-exciton
**1.850 Β± 0.001 eV** vs literature 1.85 eV. β
β
= 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
3.5β3.8 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** β 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 β `kolas-audit-predictor`
> **5-fold CV on synthetic data: accuracy 60.6% Β± 3.1pp Β· ROC-AUC 0.628 Β± 0.038
> Β· Brier 0.241 Β· F1 0.636** (n=2000 Γ 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) Β· `personnel_turnover` (0.25) Β· `sop_completeness` (0.24) β 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`](docs/HONESTY_NOTES.md) for citation rules.
---
## Trustworthy AI
MetroAI treats AI-output trust as a first-class engineering problem β full policy in
[`docs/TRUSTWORTHY_AI.md`](docs/TRUSTWORTHY_AI.md):
- **Data-origin labels** β every dataset and metric tagged real / synthetic / stub
- **Honest metrics** β publication rules in [`docs/HONESTY_NOTES.md`](docs/HONESTY_NOTES.md) (see synthetic-data caveats above)
- **Verifiable audit trail** β Ed25519 signatures + W3C PROV-O provenance graphs
- **Uncertainty quantification** β 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 β experimental `dcc` module)
- RFC 8032 (Ed25519 signatures)
---
## DCC β 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](https://wiki.dcc.ptb.de))
in a KOLAS / ISO 17025 context.
What it does today β honest scope:
- **`dcc.parser`** β 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`** β 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` β a draft with warnings is not a certificate.
- **`dcc.units`** β common unit strings β D-SI notation (`"mm"` β `\milli\metre`,
`"Β°C"` β `\degreecelsius`); unmapped units are kept verbatim with a warning.
- **`metroai/dcc/kolas_map.md`** β mapping table: KOLAS / ISO 17025 Β§7.8
certificate content requirements β DCC 3.3.0 elements, including known gaps.
```python
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 β v2-spec pages
**v2 backbone (since v0.6.0):**
1. **π Landing** (`app.py`) β KOLAS Compliance OS positioning + domain wizard
2. **π€ 6 Agents Dashboard** (`pages/11`) β Quality Manager daily view, KPI strip + task queue
3. **π SOP Gap Analyzer** (`pages/12`) β Technical Manager work surface, AI-detected gaps + **v0.7 domain-specific checklist**
4. **π° KOLAS Feed** (`pages/13`) β kolas-monitor regulatory news
5. **π― Audit Risk Detail** (`pages/14`) β explainability, waterfall + AI reasoning + what-if
6. **π
Ops Backbone** (`pages/15`) β certificates / personnel / schedule
**v0.7.0 P0 β lab-operator journey (NEW):**
7. **π¬ SEM domain dashboard** (`pages/16`) β SEM-EDS standards + KOLAS process + nonconformities + SOP checklist
8. **βοΈ TEM domain dashboard** (`pages/17`) β lattice constant, ISO 29301 + Cs-corrector spec
9. **π AFM domain dashboard** (`pages/18`) β surface roughness Sa/Sq/Sz per ISO 25178-2
10. **π OCD domain dashboard** (`pages/19`) β Scatterometry / RCWA library matching per SEMI MF-1789
11. **π KOLAS application form** (`pages/20`) β Fill-once β 7-section ISO 17025-style PDF (KAB-F-21 reference)
Plus the legacy v0.5 calibration / PT / certificate pages (`pages/1`β`10`).
---
## Repository layout
```
metroai/
βββ app.py β v2-spec landing page (Streamlit entry)
βββ pages/ β Streamlit multi-page
β βββ 1_π_λΆνλ_κ³μ°.py β Uncertainty calculator (KR)
β βββ 2_π_PT_λΆμ.py β PT analysis (KR)
β βββ 3_π_κ΅μ μ±μ μ.py β Calibration certificate PDF
β βββ 4_π_λΆνλ_μμ€κ³.py β Reverse uncertainty (novel)
β βββ 11_π€_6_Agents.py β v2 block 2: main dashboard
β βββ 12_π_SOP_κ°_λΆμ.py β v2 block 4: SOP gap analyzer
β βββ 13_π°_KOLAS_νΌλ.py β v2 block 5: regulatory feed
β βββ 14_π―_κ°μ¬_μν_μμΈ.py β v2 block 3: risk explainability
β βββ 15_π
_μΈμ¦μ_μΈλ ₯_μΌμ .py β v2 block 6: operations
βββ metroai/
β βββ core/ β GUM / MCM / model parsing
β βββ agents/ β 6 AI agents backbone
β βββ audit/ β Ed25519 + PROV-O
β βββ connectors/ β KOLAS / DART / NTIS live fetch + stub fallback
β βββ math/ β Sobol QMC
β βββ ml/ β GBT audit-risk model + synthetic data
β βββ templates/ β 9 calibration templates
β βββ dcc/ β NEW v0.8.x: DCC (digital calibration
β β certificate) parser/builder β experimental
β βββ inverse/ β NEW v0.8.0: uncertainty-aware ML inverse metrology
β β βββ __init__.py β package: cores + INSTRUMENTS map (11)
β β βββ uncertainty.py β β unified GUM core
β β βββ ml_inverse.py β β‘β’ ML inverse + ML uncertainty core
β β βββ metrology_module_2..10_*.py β 9 instrument modules
β β β (XRR/TEM lattice/Raman/TEM strain/SEM/AFM/PL/NSOM/Lamb)
β β βββ flagship_v0_forward_inverse.py β OCD forward+library inverse (R+T=1, err<2nm)
β β βββ flagship_v1_autodiff_gpu.py β OCD autodiff inverse (Meent torch)
β β βββ ocd_depth1..2p6_*.py β OCD accuracy / GPR / noise-robustness deep-dive
β β βββ nist_real_data_inverse.py β NIST L100P300 real-die inverse
β β βββ PLATFORM_INDEX.md β inverse engine map + honest status
β βββ schemas.py β Pydantic v2 input validation
β βββ exceptions.py β MetroAIError hierarchy
β βββ mcp_server.py β MCP stdio server
βββ tests/ β 243 unit tests (pytest)
βββ docs/
β βββ HONESTY_NOTES.md β Citation rules
β βββ TRUSTWORTHY_AI.md β Trust & governance policy
β βββ v0.7.0_ROADMAP.md β Next 3 months
β βββ RELEASE_NOTES_v0.6.0.md
βββ mcp_manifest.json β MCPize manifest (v0.6.0)
βββ pyproject.toml
```
---
## Roadmap (v0.7.0 β v0.8.0 β 2026-05 β 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 β user-fit-first**.
| Priority | Item | Status | Goal |
|---|---|---|---|
| **P0** | Domain-specific entry wizard (SEM/TEM/AFM/OCD/general) | β
shipped v0.7.0 | Stage 1 of journey |
| **P0** | Domain-specific KOLAS guides | β
shipped v0.7.0 | Stage 2, 4 |
| **P0** | KOLAS application form auto-generator | β
shipped v0.7.0 | Stage 3 |
| **P0** | Domain SOP rule-based checklist | β
shipped v0.7.0 | Stage 5 |
| **P0** | Inverse engine: 2 shared cores (GUM + ML uncertainty) + 11 instrument modules | β
shipped v0.8.0 | Forward U β inverse param+U |
| P1 | Inverse engine: real-data benchmark expansion (XRR / TEM / AFM measured) + pytest in CI | in progress | Lift β
β
β β
β
β
|
| P1 | Real KOLAS audit data + GBT retrain | pending | Replace synthetic 60.6% |
| P2 | HF Spaces migration | planned | Eliminate Streamlit Cloud sleep |
| P3 | Consulting SOP guide (per-domain on-site eval prep) | needs author | Cover stage 7 partially |
| P3 | LLM-assisted kolas-monitor (real inference) | stub now | Clear AI differentiation |
See [`docs/v0.7.0_ROADMAP.md`](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
```bash
pip install -e ".[dev,ml]"
pytest tests/ -v
```
Latest CI on Python 3.10 / 3.11 / 3.12 β **243 passing tests** across
the v0.5 β 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](LICENSE).
---
## Community
- **GitHub Issues** β bug reports, feature requests
- **GitHub Discussions** β Q&A, design discussion
- **MCPize page** β install + reviews: [mcpize.com/mcp/measurement-uncertainty](https://mcpize.com/mcp/measurement-uncertainty)
- **Glama listing** β [glama.ai/mcp/servers?query=metroai](https://glama.ai/mcp/servers?query=metroai)
- **Email** β kyb8801@gmail.com (KOLAS-side feedback especially welcome)
---
## νκ΅μ΄ μ¬μ©μλ₯Ό μν μλ΄
λ³Έ νλ‘μ νΈλ νκ΅ KOLAS μΈμ κΈ°κ΄ μ€λ¬΄μκ° μ§μ μ¬μ©ν μ μλλ‘ νκ΅μ΄ νμ΄μ§μ
νκ΅μ΄ UI λ₯Ό μ§μν©λλ€. μμΈν νκ΅μ΄ κ°μ΄λλ [`docs/RELEASE_NOTES_v0.6.0.md`](docs/RELEASE_NOTES_v0.6.0.md)
λ° Streamlit μ±μ νκ΅μ΄ νμ΄μ§λ€ (λΆνλ κ³μ° / PT λΆμ / κ΅μ μ±μ μ / λΆνλ μμ€κ³
/ KOLAS λ‘λ맡 / 6 Agents λμ보λ / SOP κ° λΆμ / KOLAS νΌλ / κ°μ¬ μν μμΈ /
μ΄μ λ°±λ³Έ) μ μ°Έκ³ ν΄μ£ΌμΈμ. cold-feedback νμν©λλ€ β `kyb8801@gmail.com`.
v0.8.0 λΆν°λ **μκ³μΈ‘ μμ§**(`metroai.inverse`)μ΄ μΆκ°λμμ΅λλ€ β μΈ‘μ μ νΈ(μ€ννΈλΌΒ·
μ΄λ―Έμ§Β·νμ )λ‘λΆν° νλΌλ―Έν°μ κ·Έ λΆνλλ₯Ό λμμ 볡μνλ©°, 11κ° μ₯λΉκ° κ³΅μ© GUM μ½μ΄μ
ML-λΆνλ μ½μ΄λ₯Ό ν¨κ» νΈμΆν©λλ€. ν©μ±Β·μ€μΈ‘ λ±κΈ(β
)κ³Ό νκ³(GPR λ
Έμ΄μ¦ λΆκ΄΄, ML μ μ©
λ²μ)λ₯Ό README μ μ μ§νκ² νκΈ°νμ΅λλ€.
---
> Built with care by [@kyb8801](https://github.com/kyb8801) Β· KOLAS RMP operations background.
This server cannot be deployed
Maintenance
ActivitySlowing
ResponsivenessNo issues