Mathematics
71+ deterministic maths primitives for linear algebra, statistics, and ML workflows.
Open tools for theresearch stack
Aqwel AI turns opaque ML workflows into auditable, reproducible pipelines—with open libraries, datasets, and documentation built for researchers and engineers.
pipeline.pyimport aion
m = aion.maths
cov = m.covariance(
X[0], X[1])
RESEARCH
accuracy · f1
pipeline.pyimport aion
m = aion.maths
cov = m.covariance(
X[0], X[1])
• 01
Aion
Parse, compute, and ship research. Openly.
Precision mathematics and experiment orchestration for AI researchers—stateless operators, auditable I/O, and 500+ primitives in one Python library.
v0.2.0 on PyPI
Visualization
Heatmaps, training plots, and matrix debugging visuals built on matplotlib.
Research stack
Embeddings, RAG primitives, remote providers, and a NumPy Former in one import surface.
What's new
Aion v0.2.0 vs v0.1.9
Compare the previous release with the research-library expansion—physics, universe, vision, core ML, and platform services—while keeping the v0.1.9 providers, tools, RAG, and Former stack.
| Dimension | v0.1.9 | v0.2.0 |
|---|---|---|
| Focus | Providers, tools, RAG, Former, native fast_* | + physics, universe, vision, core ML, cache, serve/db, Hub |
| Modules | ~15 documented core groups | 50+ modules / 500+ primitives |
| Science | — | physics, universe |
| Vision / ML | Visualization + evaluate | vision, preprocessing, models, metrics, hyperopt |
| Platform | CLI, config, I/O | + serve, db, store, tracker, usage, hub, doctor |
| Upgrade | pip install aqwel-aion | pip install -U "aqwel-aion>=0.2.0" |
Upgrade with pip install -U "aqwel-aion>=0.2.0". Full API reference and module docs are on the Aion documentation page.
Real code
Try v0.2.0 in a few lines
Copy-ready snippets for physics, astronomy, vision, and core ML—the same patterns documented in the API reference.
Natural-language physics query
Solve classical physics tasks with aion.physics — or run a pendulum simulation.
Open docs →python
from aion.physics import solve_physics_query, simulate_pendulum
result = solve_physics_query("kinetic energy mass=2 velocity=3")
print(result.output_value) # 9.0
sim = simulate_pendulum(1.0, 0.2, steps=100)
print(sim.summary["small_angle_period_s"])CLI & workflows
Beyond the import — doctor, Hub, CLIs
Upgrade, check your environment, launch the Hub dashboard, and explore science CLIs shipped with v0.2.0.
Upgrade to v0.2.0
Pull the latest release and optional extras for vision, serve, or the full stack.
$ pip install -U "aqwel-aion>=0.2.0"Extras: [vision] [serve] [db] [full]
Environment health check
Verify Python, numpy, optional deps, and native backends before experiments.
$ aion doctorAlso: aion welcome · aion info
Local Hub dashboard
Browse modules, dependency status, and run snippets in the browser.
$ aion startOpens http://127.0.0.1:3000
Physics & universe CLIs
Run science toolkits from the terminal without writing a notebook first.
$ aion physics --help && aion universe --helpAlso: aion vision · aion usage
Research-grade coverage across the full ML workflow
42
Manual NumPy scripts
54
SciPy + sklearn
48
Notebook-only workflows
61
Commercial MLOps
58
Generic ML utilities
72
Hugging Face ecosystem
88
Aion v0.2.0
Composite score across maths primitives, algorithms, visualization, embeddings, evaluation, docs, and reproducibility tooling — based on Aion v0.2.0 API surface.
Latest release
Aion v0.2.0
A full research stack—algorithms, visualization, embeddings, evaluation—plus physics, universe, vision, core ML, cache, serve/db, and Hub.
Rigorous primitives
71+ maths utilities, search and graph algorithms, visualization, text embeddings, and evaluation metrics—stateless, deterministic building blocks.
Science & vision
Physics and astronomy toolkits, classic computer vision on NumPy arrays, plus core ML (preprocessing, models, metrics, hyperopt).
One research stack
Providers, tool loops, RAG, cache, serve/db, tokenizer, pipelines, Hub, and a NumPy Former—one import surface with optional extras.
Install & explore
Get Aion v0.2.0
PyPI for installs, GitHub for source, docs for the full API.
$ pip install
pip install aqwel-aion
v0.2.0 on PyPI
New in v0.2.0
What's in this release
Classical physics toolkit
Mechanics, kinematics, thermo, EM, optics, relativity, integrators, and natural-language physics queries.
- Pendulum & spring sims
- NL query solver
- Optional C++ path
Astronomy toolkit
Coordinates, observing, orbits, cosmology, and catalogs with optional native acceleration.
- Sky coordinates
- Moon & observing
- Cosmology helpers
Computer vision helpers
NumPy-first image I/O, transforms, filters, draw, metrics, and OpenCV ops via [vision].
- Read/write images
- Filters & contours
- SSIM / PSNR
Core ML modules
Preprocessing, classical models, metrics, and hyperparameter search—NumPy-first.
- Scalers & encoders
- KNN / trees / PCA
- Grid & random search
Serve, store & Hub
FastAPI serve, SQLite stores, experiment tracking, usage dashboard, and local Hub.
- aion serve
- Tracker & store
- aion start / usage
Iterative & open
Aion v0.2.0 is available now. Feedback from the research community is always welcome.
View documentation →“Aion is not intended to replace existing machine learning frameworks. Its purpose is to support the research process by reducing boilerplate and improving clarity.”
Modules
Feature overview
Mathematics & Stats
71+ functions: linear algebra, statistics, ML activations and losses.
- Linear algebra
- Probability & distributions
- FFT & convolution
- Normalization & scaling
Algorithms
Search, arrays, and graph utilities for research and data pipelines.
- Binary & exponential search
- Flatten, chunk, sliding window
- BFS, DFS, A*, PageRank
Physics & Universe
Classical physics and astronomy toolkits with optional native backends.
- Mechanics & EM
- Coordinates & orbits
- Cosmology catalogs
Vision & Visualization
CV on NumPy arrays plus 1D/2D/3D plots and training metrics.
- Image I/O & filters
- Heatmaps & confusion
- 3D scatter & surfaces
AI & Evaluation
Embeddings, providers, RAG, LLM eval, prompts, and metrics.
- Text embeddings
- Tool loops & RAG
- Faithfulness & cost
Dev & Operations
Files, watcher, Git, cache, db, serve, Hub, and doctor.
- Safe I/O & cache
- aion doctor
- Hub dashboard
Why Aion
Built for research workflows
Full pipeline
From maths and algorithms to physics, vision, serve, and publication-ready docs.
71+ functions
Linear algebra, statistics, signal processing, and ML helpers.
500+ primitives
Algorithms, science modules, ML, providers, RAG, and more in one API.
Layered install
Core + numpy first; add [vision], [serve], [full], or other extras when you need them.
v0.2.0 extends the stack with physics, universe, vision, core ML, cache, serve/db, and Hub.
50+
Core modules
500+
Functions
71+
Maths
30+
Languages
• Use cases •
How researchers and teams use Aqwel AI
Reproducible experiments
Ship papers with primitives you can audit and cite.
Explore AionFree research datasets
Download curated CSV & JSON data for training and evaluation.
Browse datasetsEngineering services
Custom bots, APIs, and software when your team needs to ship fast.
View servicesDatasets
Explore & download data
88 datasets across time series, tabular, NLP, finance, sensor, healthcare, retail, and geospatial. Up to 1M rows per dataset. Free for research—JSON & CSV.
Categories
Time seriesTabularNLPFinanceSensorHealthcareRetailGeospatial
Search by name, filter by category, choose row count (10k–1M), and download as JSON or CSV. No sign-up required.
Ready for training
Wide tables with many columns—structured for ML experiments, benchmarking, and reproducible research pipelines.
Browse all datasets →Discover
New from the lab
Shorter lab notes on the Blog, and Lattice—a lattice of copy-ready prompts for research, coding, and evaluation.
Blog
Lab notes
Product updates, guides, and shorter stories from Aqwel AI—alongside deeper work on the Research Hub. New posts land here as we ship.
Visit the blog →LatticeCopy-ready prompts
Curated prompts for literature reviews, Aion workflows, datasets, writing, and evaluation. Copy once, adapt for any model.
Open Lattice →Open by default
MIT-licensed Aion, public datasets, and docs—inspect every primitive before you depend on it.
Security-conscious design
No card data on-site, strict CSP headers, and transparent open-source you can audit in GitHub.
Production ready
Built for teams running reproducible AI research with reliability, transparency, and control at every layer.
Contact usCommunity & support
GitHub issues, documentation, email support, and optional engineering services for production teams.
Flexible adoption
Install core only, add [full] or [docs] extras, or engage us for custom integrations and bots.
88+ free datasets · Aion v0.2.0 · Independent research lab in Armenia
Mission
Building for the research community
Aqwel AI exists because AI progress is often gated by opaque tooling. We ship open infrastructure so fundamental discovery stays inspectable, reproducible, and accessible.
Open tooling
Stateless maths, explicit I/O, and auditable primitives—libraries you can cite in a paper without five vendor dashboards.
Transparency first
Clear computation paths, documented optional accelerators, and APIs designed for inspection—not black-box convenience.
Reproducible workflows
From seeds and checksums to layered config—research steps that stay traceable from raw data to published claim.
We bias toward small, composable surfaces—not to compete on industrial throughput, but to make honest research workflows easier to ship and audit.
Research
Technical notes from the lab
Mechanics-first articles on machine learning, computer vision, and reproducible experimentation—written for engineers who want clarity, not hype.
LLM EngineeringJun 8, 202614 min read
Comparing Frontier LLMs: Gemini 2.5 Pro, GPT-4.1, and Claude Sonnet 4
A production-oriented comparison of Google, OpenAI, and Anthropic flagship APIs—context, modalities, tools, cost, and when to route each model in a multi-provider stack.
Read article →Computer VisionMay 9, 202620 min read
Inside CNNs: Understanding Computer Vision Through Matrix Operations
A matrix-first tour of digital images, convolution, feature maps, and pooling—built for engineers who want mechanics, not hype.
Read article →Collaborators
Partners & sponsors
Strategic partnerships that help us ship research, education, and AI-native products together.
Community voices
What builders say about our work
Feedback from researchers, engineers, and educators using Aion, our datasets, docs, and research notes in the wild.
Aion
I swapped three small NumPy helper scripts for Aion maths utilities in our lab notebook. Same outputs, fewer lines, and I could actually explain every step in the paper appendix.
Aion
The evaluate module saved us a weekend—accuracy, F1, and ROC-AUC in one place without wiring sklearn boilerplate for every ablation run.
Datasets
Wide synthetic tables with predictable seeds are perfect for teaching feature engineering. Students download CSV, preview rows in-browser, and reproduce the same splits every time.
Docs
The Aion docs site is what PyPI pages wish they were—searchable functions, copy-paste examples, and module boundaries that match the import paths.
Research
The CNN matrices article finally connected convolutions to linear algebra the way I needed for a internal tech talk. No hand-wavy diagrams—just mechanics.
Aion
Core install is lightweight; we added [full] only on the training machine. That layered dependency story is rare and honest.
Datasets
We benchmark pipeline throughput against the 100k-row exports before touching production data. The preview table alone prevented two bad schema assumptions.
Aion
Matrix heatmaps and training history plots from one library beat juggling matplotlib snippets across five repos. Good defaults, still inspectable.
Aion
I swapped three small NumPy helper scripts for Aion maths utilities in our lab notebook. Same outputs, fewer lines, and I could actually explain every step in the paper appendix.
Aion
The evaluate module saved us a weekend—accuracy, F1, and ROC-AUC in one place without wiring sklearn boilerplate for every ablation run.
Datasets
Wide synthetic tables with predictable seeds are perfect for teaching feature engineering. Students download CSV, preview rows in-browser, and reproduce the same splits every time.
Docs
The Aion docs site is what PyPI pages wish they were—searchable functions, copy-paste examples, and module boundaries that match the import paths.
Research
The CNN matrices article finally connected convolutions to linear algebra the way I needed for a internal tech talk. No hand-wavy diagrams—just mechanics.
Aion
Core install is lightweight; we added [full] only on the training machine. That layered dependency story is rare and honest.
Datasets
We benchmark pipeline throughput against the 100k-row exports before touching production data. The preview table alone prevented two bad schema assumptions.
Aion
Matrix heatmaps and training history plots from one library beat juggling matplotlib snippets across five repos. Good defaults, still inspectable.
Aion
I swapped three small NumPy helper scripts for Aion maths utilities in our lab notebook. Same outputs, fewer lines, and I could actually explain every step in the paper appendix.
Aion
The evaluate module saved us a weekend—accuracy, F1, and ROC-AUC in one place without wiring sklearn boilerplate for every ablation run.
Datasets
Wide synthetic tables with predictable seeds are perfect for teaching feature engineering. Students download CSV, preview rows in-browser, and reproduce the same splits every time.
Docs
The Aion docs site is what PyPI pages wish they were—searchable functions, copy-paste examples, and module boundaries that match the import paths.
Research
The CNN matrices article finally connected convolutions to linear algebra the way I needed for a internal tech talk. No hand-wavy diagrams—just mechanics.
Aion
Core install is lightweight; we added [full] only on the training machine. That layered dependency story is rare and honest.
Datasets
We benchmark pipeline throughput against the 100k-row exports before touching production data. The preview table alone prevented two bad schema assumptions.
Aion
Matrix heatmaps and training history plots from one library beat juggling matplotlib snippets across five repos. Good defaults, still inspectable.
Aion
Matrix heatmaps and training history plots from one library beat juggling matplotlib snippets across five repos. Good defaults, still inspectable.
Datasets
We benchmark pipeline throughput against the 100k-row exports before touching production data. The preview table alone prevented two bad schema assumptions.
Aion
Core install is lightweight; we added [full] only on the training machine. That layered dependency story is rare and honest.
Research
The CNN matrices article finally connected convolutions to linear algebra the way I needed for a internal tech talk. No hand-wavy diagrams—just mechanics.
Docs
The Aion docs site is what PyPI pages wish they were—searchable functions, copy-paste examples, and module boundaries that match the import paths.
Datasets
Wide synthetic tables with predictable seeds are perfect for teaching feature engineering. Students download CSV, preview rows in-browser, and reproduce the same splits every time.
Aion
The evaluate module saved us a weekend—accuracy, F1, and ROC-AUC in one place without wiring sklearn boilerplate for every ablation run.
Aion
I swapped three small NumPy helper scripts for Aion maths utilities in our lab notebook. Same outputs, fewer lines, and I could actually explain every step in the paper appendix.
Aion
Matrix heatmaps and training history plots from one library beat juggling matplotlib snippets across five repos. Good defaults, still inspectable.
Datasets
We benchmark pipeline throughput against the 100k-row exports before touching production data. The preview table alone prevented two bad schema assumptions.
Aion
Core install is lightweight; we added [full] only on the training machine. That layered dependency story is rare and honest.
Research
The CNN matrices article finally connected convolutions to linear algebra the way I needed for a internal tech talk. No hand-wavy diagrams—just mechanics.
Docs
The Aion docs site is what PyPI pages wish they were—searchable functions, copy-paste examples, and module boundaries that match the import paths.
Datasets
Wide synthetic tables with predictable seeds are perfect for teaching feature engineering. Students download CSV, preview rows in-browser, and reproduce the same splits every time.
Aion
The evaluate module saved us a weekend—accuracy, F1, and ROC-AUC in one place without wiring sklearn boilerplate for every ablation run.
Aion
I swapped three small NumPy helper scripts for Aion maths utilities in our lab notebook. Same outputs, fewer lines, and I could actually explain every step in the paper appendix.
Aion
Matrix heatmaps and training history plots from one library beat juggling matplotlib snippets across five repos. Good defaults, still inspectable.
Datasets
We benchmark pipeline throughput against the 100k-row exports before touching production data. The preview table alone prevented two bad schema assumptions.
Aion
Core install is lightweight; we added [full] only on the training machine. That layered dependency story is rare and honest.
Research
The CNN matrices article finally connected convolutions to linear algebra the way I needed for a internal tech talk. No hand-wavy diagrams—just mechanics.
Docs
The Aion docs site is what PyPI pages wish they were—searchable functions, copy-paste examples, and module boundaries that match the import paths.
Datasets
Wide synthetic tables with predictable seeds are perfect for teaching feature engineering. Students download CSV, preview rows in-browser, and reproduce the same splits every time.
Aion
The evaluate module saved us a weekend—accuracy, F1, and ROC-AUC in one place without wiring sklearn boilerplate for every ablation run.
Aion
I swapped three small NumPy helper scripts for Aion maths utilities in our lab notebook. Same outputs, fewer lines, and I could actually explain every step in the paper appendix.
Help center
Questions, answered
Six topics per page—search or jump between pages below.