meerax

Shared ML utilities — LLM providers, evaluation metrics, visualization, and report generation. Used as an in-house dependency across all of Sameer Maurya's ML projects and organization. Source repo: the-forge — kept its original name; only the installable package was renamed to meerax.

CI gates on mypy (strict = true, with 2 narrowly-scoped # type: ignore exceptions for known third-party stub gaps) and a minimum 85% test coverage — both enforced on every PR, not just checked locally.

Supports Python 3.9+ (requires-python = ">=3.9"). CI itself runs and is verified against Python 3.12 only — no version matrix — but installs and the full test suite have been manually verified clean on 3.9, 3.10, 3.11, and 3.12 before each release that touches this floor.

Install

pip install meerax

Or pin in requirements.txt:

meerax==1.10.0

Modules

Scaffolding Projects

Every project in the ecosystem follows the same PROJECT_STANDARDS.md layout and depends on meerax. The meerax CLI (installed alongside the package) generates or retrofits that layout:

# brand-new project
meerax new my-project --path ~/dev

# retrofit an existing, non-empty directory — additive only, never overwrites
cd ~/dev/my-existing-notebook-project
meerax init

meerax new creates the full src/{core,providers,services,utils,data} + tests/ + CI skeleton, pins requirements.txt to the current meerax release, and runs git init. The generated ci.yml calls this repo's reusable CI workflow instead of embedding its own copy, so fixes to the shared CI logic reach every project that uses it without needing to be manually reapplied. Also generates .github/dependabot.yml (pip + github-actions, weekly) so dependency pins don't quietly go stale.

Both new and init accept --template llm-report, which adds a real, runnable example on top of the bare skeleton — the "call an LLM, build an HTML report" shape that's shown up twice already (trend-whisperer, pixel-drift). It's genuinely functional code with real tests, not stub methods to fill in:

meerax new my-app --template llm-report
cd my-app && pip install -r requirements.txt
python -m src.app --prompt "Summarize this quarter's churn" --llm-provider ollama

meerax init fills in whatever's missing from that same layout without touching files that already exist, and reports any top-level files it doesn't recognize (e.g. notebooks) so you can move them into src/ by hand.

meerax doctor checks an existing project against PROJECT_STANDARDS.md — no Python version matrix, no committed docs/specs/docs/plans, a LICENSE file, VERSION/README/CHANGELOG consistency, Dependabot configured, and whether the project's meerax pin is current:

cd ~/dev/my-project
meerax doctor

Exits non-zero if anything fails, so it's safe to run in CI.

meerax bump <version> updates VERSION, the README version badge, and inserts a dated CHANGELOG heading in one step — the exact multi-file edit that's caused version-badge drift more than once across this ecosystem's history:

meerax bump 1.2.0

It doesn't write CHANGELOG content or a compare-link footer — those need someone who actually knows what changed.

Quick Start

from meerax.llm import ClaudeProvider, PromptTemplate
from meerax.eval import evaluate_classifier
from meerax.viz import apply_meerax_theme
from meerax.report import ReportBuilder, ReportSection

# LLM: swap provider without changing downstream code
llm = ClaudeProvider()                         # or OpenAIProvider() / OllamaProvider()
tpl = PromptTemplate("Explain {finding} to a risk manager in 3 sentences.")
response = llm.generate(tpl.render(finding="high AUC-ROC with low recall"))

# Eval
metrics = evaluate_classifier(y_true, y_pred, y_prob=probabilities)
print(metrics)
# Accuracy : 0.9823
# F1       : 0.8741
# AUC-ROC  : 0.9912

# Viz + Report
apply_meerax_theme()
rb = ReportBuilder("Fraud Detection — Model Report v0.1.0")
rb.add_section(ReportSection(
    title="Performance",
    metrics=metrics.to_dict(),
    content=response.content,
))
rb.save("reports/model_report.html")

LLM Provider Interface

All providers implement LLMProvider.generate() and .chat(). Swap with one line:

from meerax.llm import ClaudeProvider, OpenAIProvider, OllamaProvider

llm = ClaudeProvider()    # needs ANTHROPIC_API_KEY
llm = OpenAIProvider()    # needs OPENAI_API_KEY
llm = OllamaProvider()    # needs Ollama running locally

Benchmarks

Self-contained benchmark scripts in benchmarks/:

python benchmarks/meerax_benchmark.py            # print results
python benchmarks/meerax_benchmark.py --record   # also append to benchmarks/results/history.jsonl

--record appends one JSON line per run, keyed by the installed meerax version, so performance can be compared release to release. Run it as part of cutting a release, alongside meerax bump, and commit the updated history.jsonl in the same commit. A GitHub Actions workflow (.github/workflows/benchmarks.yml) also runs it on demand (workflow_dispatch) and uploads the results as a build artifact — informational only, not a required CI check, since wall-clock timings on shared runners are too noisy to gate on.

Project Structure

the-forge/
├── meerax/              # Installable package
│   ├── llm/            # LLM provider abstraction
│   ├── eval/            # Evaluation metrics
│   ├── viz/             # Visualization utilities
│   ├── data/            # Data loading, splitting, resampling
│   ├── report/          # HTML report builder
│   ├── scaffold/        # Project skeleton templates + create/retrofit logic
│   ├── cli.py           # `meerax new` / `init` / `doctor` / `bump` command entry point
│   ├── doctor.py        # PROJECT_STANDARDS.md compliance checks
│   ├── release.py       # VERSION/README badge/CHANGELOG bump helper
│   ├── vision/          # Image dataset loader + translation-grid plotting
│   └── logging.py       # Structured logger
├── benchmarks/          # meerax's own perf benchmarks + standalone ML demo scripts
│   └── results/         # history.jsonl — one line per --record run, tracked across releases
├── tests/
│   ├── unit/            # 152 unit tests, zero external deps
│   └── integration/     # 4 cross-module pipeline tests
├── ARCHITECTURE.md
├── SECURITY.md
├── LICENSE
├── pyproject.toml
├── requirements.txt
└── VERSION

sameer-portfolio · mauryasameer.com