GitHub - Siddardth7/quality-platform: Integrated manufacturing quality platform — FMEA, SPC, and Control Plan tools over a shared core (AIAG/IATF-16949 core tools).

Quickstart · Tools · Hosts · The Loop · Docs

Verified AIAG core-tool engines — FMEA, SPC, Control Plan and MSA / Gage R&R —
callable by hand in a browser, or by any AI agent through an MCP server and Agent Skills.

🔎 What this is

In real quality departments, the AIAG / IATF-16949 core tools live in disconnected spreadsheets and one-off apps. A failure mode found in an FMEA never becomes a control on a Control Plan, and an out-of-control point on an SPC chart never makes it back to the FMEA's risk rating. The methodology describes a closed loop; the tooling almost never implements one.

Quality Platform builds that loop for real — credible standalone tools first, everything they share promoted into a single typed core (quality_core), then wired into an end-to-end workflow you run on demand over a project's files. The last leg is deliberately not automatic: SPC evidence proposes an occurrence rating for a human to review, it never rewrites one. Proven on real semiconductor process data.

_{🛡️ FMEA Risk Analyzer — RPN & AIAG-VDA Action Priority, risk-tier triage, auto-generated insight.}	_{📈 SPC Control Charts — X̄-R / I-MR / c-charts with Western Electric & Nelson rule overlays.}
_{📊 Process Capability — Cp/Cpk/Pp/Ppk with a stability gate: no capability claim on an out-of-control process.}	_{🏭 Unified shell — every tool under one st.navigation surface, one theme, one URL.}

▶ Open the live demo →

⚖️ What it is / what it is not

It is	It is not
On-demand analysis over files you supply — every result comes from a call you or your agent makes	Not real-time or 24/7 monitoring. Nothing polls, streams or watches a line
A set of standards-anchored engines, every constant cited in an `ASSUMPTIONS_LOG.md`	Not a data historian and not an EQMS — no database, no equipment integration, no retention or e-signature layer
A loop that proposes an occurrence rating from SPC evidence, with an action pointing at the evidence file	Never an autowrite. `Cause.occurrence` is never overwritten — a human reviews the proposal and decides
Local by default — the MCP server's stdio transport runs on your machine, so your data never leaves it	Not a replacement for engineering judgment. The tools compute, cite and refuse; the engineer decides

Every one of those lines, with the reasoning behind it, is on the docs site: Limitations.

🔌 Why MCP-first — the idea

Quality work happens where the engineer already is. So the engines ship as MCP tools and Agent Skills first, and as a UI second — the same quality_core code underneath both, so the two cannot drift apart.

Local, no database, no upload. stdio is the default transport: your agent host launches the server as a local process and talks to it over pipes. There is no port, no token and no service — the data stays on the machine that owns it.
One config, any host. 49 tools behind a single stdio config block that four named hosts accept nearly verbatim. Which hosts, which quirks, and — honestly — which configurations have actually been run rather than merely written: Hosts.
The agent does not do the arithmetic. The point of the tool layer is that fmea_score(8, 5, 6) returns {"rpn": 240, "action_priority": "Medium"} from tested, coverage-gated code — not from a model's head.
Roadmap. MCP + skills today. A standalone status-product website is a Phase-2 item that is not built and not live; nothing on this page depends on it.

🚀 Quickstart

Clone first — nothing is published to a package index yet (uvx quality-mcp does not work; #292 is open).

git clone https://github.com/Siddardth7/quality-platform.git
cd quality-platform
uv sync

1 · As MCP tools in your agent host

uv run python -m mcp_app.server    # from the workspace root

Register it — this is the Claude Desktop shape (root key mcpServers); Cursor and Gemini CLI take the same block in their own config files, and VS Code uses servers instead:

{
  "mcpServers": {
    "quality-platform": {
      "command": "uv",
      "args": [
        "run",
        "--directory", "/absolute/path/to/quality-platform",
        "python", "-m", "mcp_app.server"
      ]
    }
  }
}

Then ask your agent: "Score this failure mode: severity 8, occurrence 5, detection 6" — the host should call fmea_score and report the tool's result, rpn 240 / "Medium". health() and version() are the cheaper liveness checks. Run the whole loop with run_project_loop("examples/secom-quality-loop"), or a capability study with spc_capability.

Per-host config blocks, the quirks table and the PASS/PENDING record: Hosts → apps/mcp/docs/HOSTS.md.

2 · As a local Streamlit app

uv run streamlit run app.py            # the whole platform — one URL, every tool
uv run streamlit run apps/spc/app.py   # a single app standalone

Run a single app from its own directory

cd apps/fmea && streamlit run app.py   # FMEA Risk Analyzer
cd apps/spc  && streamlit run app.py   # SPC Dashboard

Each app still runs unchanged from its own directory.

3 · The worked loop

uv run python -c "from mcp_app.server import run_project_loop; \
    run_project_loop('examples/secom-quality-loop')"

The write-up — including which files in it are real SECOM data and which are illustrative — is at docs/demo.md and examples/secom-quality-loop/.

🔄 The loop

The architectural payoff: the AIAG core-tools loop, wired end to end and run on real data. It runs on demand over a project directory on disk — one call, four arrows — not continuously.

flowchart LR
    FMEA["🛡️ FMEA<br/>score S·O·D →<br/>RPN / Action Priority"]
    CP["🧩 Control Plan<br/>failure mode → characteristic,<br/>spec, method, sample plan,<br/>recommended chart"]
    SPC["📈 SPC<br/>control charts +<br/>capability (Cp/Cpk)"]
    MSA["📏 MSA / Gage RR<br/>is the measurement<br/>system even trustworthy?"]
    SECOM[("🏭 SECOM<br/>real semiconductor<br/>process data")]

    FMEA -->|"high-risk items<br/>become controls"| CP
    CP -->|"auto-configures<br/>the chart"| SPC
    SPC -->|"proposed occurrence-rating /<br/>CAPA (human reviews)"| FMEA
    MSA -.->|"prove the gage<br/>before trusting the chart"| SPC
    SECOM -.->|"runs through<br/>every tool"| SPC

    classDef live fill:#0b1220,stroke:#e65100,stroke-width:2px,color:#fff;
    class FMEA,SPC,CP,MSA,SECOM live;

Two things the diagram cannot say by itself:

spc/results/*.json is an input, not an output. No arrow writes it — it is the trace of a prior, separate charting session, read off disk as a precondition. With none there, the feedback leg legally no-ops and returns null.
The feedback arrow proposes. It attaches a candidate Action whose owner points at the evidence file and leaves Cause.occurrence untouched — "it NEVER writes a new rating, it only proposes one for a human to review." Re-running with unchanged inputs rewrites nothing.

Full detail: The loop.

🧰 The tools

Tool	What it does	Status
🛡️ FMEA Risk Analyzer	Failure Mode & Effects Analysis — RPN + AIAG-VDA Action Priority, editable S/O/D scales, relational model (Function → FM → Effect / Cause / Control), action tracking, Pareto + risk heatmap, Excel/PDF/CSV export
📈 SPC Dashboard	Statistical Process Control — variables & attributes control charts, Western Electric / Nelson rules, Cp/Cpk/Pp/Ppk with a stability gate, live disturbance simulator
🧩 Control Plan connector	Turns FMEA failure modes into a Control Plan (characteristic, spec, method, sample plan, recommended chart) — the APQP-adjacent bridge that closes the loop
📏 MSA / Gage R&R	Measurement Systems Analysis — Gage R&R (Average-and-Range by default, or ANOVA with the part×appraiser interaction), %EV/%AV/%GRR/%PV vs study & tolerance, `ndc`, accept/marginal/reject vs AIAG thresholds
🔌 MCP server	49 tools over the same engines — Meta (2) · FMEA (5) · SPC (19) · project-file arrows & the loop (4) · export/report (13) · Control Plan (4) · MSA (1) · private-corpus RAG (1). Full grouped catalog: tool catalog
📚 `quality-research` skill	Answers "why does the standard say X" by calling `qdb_answer_question` for a cited, page-located answer, falling back to this repo's public `ASSUMPTIONS_LOG.md` citations when the endpoint is unreachable. Explicitly not for running an engine — that is the `fmea` / `spc` / `msa` / `control-plan` skills
🏭 SECOM case study	The whole platform run on real semiconductor sensor data — SPC, yield/DPPM, Pareto of failing signals. No Cp/Cpk and no Gage R&R: the dataset structurally supports neither, and both are refused rather than invented

Standards context: FMEA — AIAG-VDA (2019) + AIAG FMEA-4 · SPC — AIAG SPC 4th Ed. · capability target Cpk ≥ 1.33. The AIAG-VDA Action Priority table is verified cell-by-cell against the primary handbook.

🔗 Hosts

Six hosts have a written, schema-correct MCP-server config in apps/mcp/docs/HOSTS.md. What has actually been run is tracked separately from what has been written, on purpose:

Host	Transport	Config verified	Worked example run
Claude Desktop · Cursor · VS Code	stdio	PASS ✓ (shape)	PENDING — GUI apps, not launchable headless
Gemini CLI	stdio	PASS ✓ — the host spawned the server and enumerated its tools	PENDING — the host's non-interactive permission gate auto-denied the call
Claude.ai · ChatGPT	http	PASS ✓ (shape only)	PENDING — no hosted endpoint and an OAuth-vs-bearer gap (#355)
Claude Code · Codex CLI · Cursor · Gemini CLI (skill layer)	stdio via skill scripts	—	PASS ✓ on all four — `skills/COMPATIBILITY.md`

A PASS ✓ in the config column is not permission to say the host works. The last row is the skill-script layer, which is a different layer from native MCP-server registration — the two do not overlap. HTTP transport is opt-in (MCP_TRANSPORT=http), always authenticated, and fails closed. The registry manifests (server.json, smithery.yaml, glama.json) exist but none is live-published.

🎓 Standards & fidelity

The public validation story, in three checkable pieces:

Every constant is cited. Each app carries docs/ASSUMPTIONS_LOG.md listing every AIAG/ISO constant, threshold and quotation with its source. A value cannot change without its log changing.
MSA's citations are machine-checked. apps/msa/docs/CITATIONS.tsv is a manifest asserted in CI by apps/msa/tests/test_citations.py — a drifted or fabricated quotation fails the build, not a review.
Nine coverage gates at 100% with branch coverage on (see below). Where no published standard exists, the module says so in its own docstring rather than implying one.

Refusals count too: no capability claim on an out-of-control process, no Gage R&R on SECOM (it has no part/appraiser/trial axis), no Cp/Cpk on SECOM (it ships no tolerances). More: Standards & fidelity.

🏗️ Architecture

A uv workspace monorepo: four Streamlit apps mounted under one shell, SECOM as an engine-only member (a tested library, deliberately not mounted), and an MCP server exposing the same engines as tools. Every cross-cutting concern is written once in quality_core and consumed by all of them.

flowchart TB
    subgraph Agents["🔌 Agent hosts · MCP + Agent Skills"]
        MCP["apps/mcp<br/>FastMCP server · 49 tools"]
    end
    subgraph Shell["🏭 Unified shell · app.py (st.navigation)"]
        Home["Landing + one theme + one nav"]
    end
    subgraph Apps["Apps · mounted in the shell"]
        FMEA["🛡️ FMEA<br/>apps/fmea"]
        SPC["📈 SPC<br/>apps/spc"]
        CP["🧩 Control Plan<br/>apps/controlplan"]
        MSA["📏 MSA<br/>apps/msa"]
    end
    subgraph Engine["Engine-only · library, not mounted"]
        SECOM["🏭 SECOM<br/>apps/secom"]:::engine
    end
    subgraph Core["📦 packages/quality-core → import quality_core"]
        Schema["schema/<br/>flat + relational contracts (Pydantic v2)"]
        IO["io/<br/>validated ingest · CSV/Excel/PDF export"]
        Scoring["scoring.py<br/>RPN · AIAG-VDA Action Priority"]
        Theme["theme/<br/>palette · style"]
    end

    Home --> FMEA & SPC & CP & MSA
    MCP --> FMEA & SPC & CP & MSA
    FMEA --> Schema & IO & Scoring & Theme
    SPC --> IO & Theme
    CP --> Schema & IO & Scoring
    MSA --> Schema & IO & Theme
    SECOM --> IO
    SECOM -.reuses the SPC engine.-> SPC

    classDef engine opacity:0.85,stroke-dasharray:4 4;

Why it's built this way

Shared core, consumed many times. quality_core.io owns CSV/Excel/PDF export (formula-injection safe) and validated ingest — so upload validation and export are guaranteed identical across tools, and across the UI and the MCP tools. That's the economic argument of a monorepo, made concrete and coverage-gated at 100%.
Schema promoted only when stable. Contracts lived inside the FMEA app until they earned promotion to quality_core.schema — deferred extraction, done once, correctly.
History preserved. The FMEA and SPC apps were previously standalone repos, migrated here with full commit history intact — the histories are part of the engineering story.

🛡️ The quality gate

The whole workspace shares one quality bar (ruff.toml, mypy.ini, pytest config in pyproject.toml). It runs locally and, identically, in CI on every push and PR to main — a protected branch that requires the gate to pass before merge.

uv run ruff check .     # lint + format check
uv run mypy             # strict static types
uv run pytest --cov     # 2502 tests + coverage across core + apps

Coverage gates — CI-enforced, cannot silently regress:

Surface	Bar
`quality_core.io` — shared export + ingest	100%
`quality_core.schema` — shared FMEA contracts	100% (line + branch)
SPC testable surface — engine + simulation + visualizer + exporter	≥ 95%

Workflow discipline: one logical change per commit (conventional commits) · one issue at a time · multi-agent code review before finishing · push → CI green → close issue → tag a release each week · if a week can't ship green, cut scope, not quality.

🗺️ Roadmap

Twelve tracked weeks, one release each, ending on a portfolio-grade v1.0.0.

Phase	Weeks	Focus
A · Foundation	1–2	Monorepo, shared core, shell, one CI gate · `v0.1–v0.2` ✅
B · Standards-correct cores	3–5	AP-native + relational FMEA, shared validation/export · `v0.3–v0.5` ✅
C · Integration & core-tool completion	6–9	Control Plan → close the loop → MSA → SECOM real-data case study
D · Depth & legibility	10–12	Modern SPC depth, DOE on SECOM, then a hardening pass → `v1.0.0-portfolio`

_{An explainable AI FMEA copilot (LLM + RAG + eval harness) is a documented, unscheduled future
phase. The full plan — vision, diagrams, week-by-week detail — lives in ROADMAP.md.}

📁 Repository layout

quality-platform/
├── app.py                  # unified platform shell (st.navigation)
├── shell/                  # landing page + shared chrome
├── mkdocs.yml              # docs site config (MkDocs Material)
├── docs/                   # docs-site pages + engineering docs
├── skills/                 # Agent Skills — fmea · spc · msa · control-plan · project-loop · quality-research
├── examples/               # worked examples — secom-quality-loop/
├── ROADMAP.md              # the full project guide (vision, diagrams, 12-week plan)
├── packages/
│   └── quality-core/       # shared core  →  import quality_core
│       └── src/quality_core/
│           ├── schema/     # flat (FMEARow) + relational (Function→FM→…) contracts
│           ├── io/         # validated ingest · CSV/Excel/PDF export (injection-safe)
│           ├── scoring.py  # RPN · AIAG-VDA Action Priority
│           └── theme/      # palette · style
└── apps/
    ├── fmea/               # FMEA Risk Analyzer  (full original history preserved)
    ├── spc/                # Manufacturing SPC Dashboard  (full original history preserved)
    ├── controlplan/        # Control Plan connector — FMEA → characteristic/spec/method/chart
    ├── msa/                # MSA / Gage R&R — Average-and-Range (default) + ANOVA
    ├── secom/              # SECOM real-data case study — engine-only, not mounted in the shell
    └── mcp/                # MCP server — the engines as 49 tools (FastMCP, stdio)

_{Migrated from the standalone repos
fmea-risk-analyzer and
manufacturing-spc-dashboard,
now archived → moved here.}

📚 Documentation

The full docs site — quickstart, per-engine pages, the 49-tool catalog, the host matrix, the worked example, standards and limitations — is at:

→ siddardth7.github.io/quality-platform

In this repo	What it is
`docs/demo.md`	The SECOM worked example, end to end
`docs/tools.md`	The 49 MCP tools, grouped by method
`docs/limitations.md`	What this platform deliberately does not do
`apps/mcp/docs/HOSTS.md`	Per-host MCP configuration + PASS/PENDING record
`docs/DEFINITION_OF_DONE.md`	The contract every change is held to

🧱 Built with

New here? Start with the docs site or the ROADMAP.md · Contributing? See CONTRIBUTING.md and the Definition of Done.

_{Manufacturing-quality engineering, built like software — typed, tested, and shipped weekly.}

Navigation Menu

Repository navigation

About

Repository files navigation

🔎 What this is

⚖️ What it is / what it is not

🔌 Why MCP-first — the idea

🚀 Quickstart

1 · As MCP tools in your agent host

2 · As a local Streamlit app

3 · The worked loop

🔄 The loop

🧰 The tools

🔗 Hosts

🎓 Standards & fidelity

🏗️ Architecture

🛡️ The quality gate

🗺️ Roadmap

📁 Repository layout

📚 Documentation

🧱 Built with

About

Releases

Packages

Contributors

Languages

Name		Name	Last commit message	Last commit date
.agents		.agents
.claude		.claude
.devcontainer		.devcontainer
.github/workflows		.github/workflows
apps		apps
assets		assets
docs		docs
examples/secom-quality-loop		examples/secom-quality-loop
packages/quality-core		packages/quality-core
scripts		scripts
shell		shell
skills		skills
tests		tests
.gitattributes		.gitattributes
.gitignore		.gitignore
.python-version		.python-version
CHANGELOG.md		CHANGELOG.md
CLAUDE.md		CLAUDE.md
CONTRIBUTING.md		CONTRIBUTING.md
README.md		README.md
ROADMAP.md		ROADMAP.md
app.py		app.py
glama.json		glama.json
mkdocs.yml		mkdocs.yml
mypy.ini		mypy.ini
pyproject.toml		pyproject.toml
requirements.txt		requirements.txt
ruff.toml		ruff.toml
uv.lock		uv.lock

Latest commit

History

Folders and files

Repository files navigation

🔎 What this is

⚖️ What it is / what it is not

🔌 Why MCP-first — the idea

🚀 Quickstart

1 · As MCP tools in your agent host

2 · As a local Streamlit app

3 · The worked loop

🔄 The loop

🧰 The tools

🔗 Hosts

🎓 Standards & fidelity

🏗️ Architecture

🛡️ The quality gate

🗺️ Roadmap

📁 Repository layout

📚 Documentation

🧱 Built with

About

Resources

Contributing

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages