Raw test-stand data becomes ready-to-use engineering analytics — in minutes. Built and validated on real hot fire tests, on live telemetry.
From the acquisition export to a complete descriptive and diagnostic report — with no manual preparation in between.
Sixteen tabs of descriptive analysis, ten of diagnostics, eight campaign sections — per test, every time.
Performance computed from your own measurements, with geometry and assumptions stated beside every result.
Built and proven alongside live hot-fire campaigns on a working test stand — never on synthetic data.
It explains the computed result in plain language — and never produces the number itself.
Aexion Synthesis builds the data analytics infrastructure for liquid rocket engine hot fire testing. We turn raw test-stand telemetry into structured, engineering-grade analytics — performance, diagnostics, anomaly detection — within minutes of a test, with an AI assistant that answers questions in plain language and always traces back to the source data.
The platform was built and proven on real hot fire campaigns, on live telemetry from a working test stand — not synthetic data. Our team does propulsion engineering itself, which is why the analytics speak the language of the people who run the tests. Founded in 2024 — and already validated on live hot fire campaigns, backed by NVIDIA Inception and the AWS and Anthropic startup programs. Remote-first, based in Almaty, Kazakhstan.
The cloud foundation of the platform. Ingest, computation, storage, and the portal all run on AWS infrastructure. AWS supports the project with credits and hands-on architecture guidance — helping us build a platform that scales from a single test to full campaigns without compromising data security.
aws.amazon.com ↗Member of the global NVIDIA program for AI companies: access to NVIDIA technology and engineering expertise, training resources, and an ecosystem that includes leading aerospace startups — a direct line to the tooling and practices behind modern AI-driven engineering analytics.
nvidia.com ↗Member of the Claude for Startups global program. Claude is the foundation of the platform's AI analyst — the layer that explains computed results in plain language, in English, Russian, and Kazakh, while every number stays with the deterministic computation core.
claude.com/startups ↗Our test-stand partner. Thrust designs and fires its own liquid rocket engines and holds a state licence for space activity — the first private company in Kazakhstan to do so. The platform was built alongside their hot fire campaigns and proven on their telemetry: real hardware, real stand, not synthetic data.
thrust.kz ↗One of the region's largest distributors of IT solutions and cloud services, including AWS and Azure. MUK is our procurement channel for the region: cloud infrastructure and vendor licensing behind platform deployments across Kazakhstan and Central Asia.
muk.group ↗Our cloud engineering partner — an AWS Advanced Tier partner with the AWS AI competency and a member of the AWS Well-Architected Framework program, with a certified engineering team building and operating AWS solutions since 2021, backed by 24/7 managed services. qCloudy supports the platform's cloud infrastructure, architecture reviews, and operations.
qcloudy.io ↗Wherever liquid rocket engines are fired, someone has to turn the data into an answer — usually before the next test is scheduled. That is the job Aexion is built for.
Turn every hot-fire test into a finished engineering record — the same day.
A single firing leaves gigabytes across dozens of channels, and the read usually waits days for someone to have time. Performance, ignition classification and flagged deviations are computed before the stand is cold.
Engineering-grade analytics without standing up a data team.
You need engines qualified, not a data pipeline maintained. Derived performance, stability metrics and deviation review from the first test — without hiring for it or writing throwaway analysis scripts.
Reproducible, traceable results across an entire campaign.
Work has to hold up years later, reviewed by people who were not on the stand. Every test in a campaign runs through the same deterministic core, and every number stays traceable to the channel and time window it came from.
Hanabi is the analytics core behind every Aexion report. It takes the raw output of your acquisition system — thrust, chamber and feed pressures, propellant temperatures, mass flow, valve states, stand video — and returns a complete engineering read of the firing: what the engine did, how well it did it, and what needs a closer look before the next test.
Export from your acquisition system as it was written. Channel names are mapped once per stand, then reused for every test after that.
Derived parameters, phase statistics, ignition classification and deviation flags — produced by a deterministic core, not estimated.
An interactive report per test, plus exportable documents for design reviews and programme records.
| IGNITION TRANSIENT | t+0…2.1 s | peak thrust 428 kgf |
|---|---|---|
| STEADY-STATE WINDOW | t+2.1…9.6 s | mean 352 kgf · CV 3.1% |
| SHUTDOWN | t+9.6…10.4 s | tail-off measured, not trimmed off |
| CHANNELS | THRUST-01, PC-01, PF-02, MF-01/02 | one time base, one clock |
| STAND VIDEO | CAM 01 | locked to the same clock |
Statistics are computed inside each phase and never averaged across them: a transient and a steady burn are different measurements of different things.
| CHARACTERISTIC VELOCITY | 1471 m/s | c* = Pc · A* / ṁ |
|---|---|---|
| EFFICIENCY | η_c* 0.90 | measured against a thermochemical baseline |
| THRUST COEFFICIENT | CF 1.42 | CF = F / (Pc · A*) |
| SPECIFIC IMPULSE | 213 s | Isp = c* · CF / g₀ |
| MIXTURE RATIO | O/F 1.35 | from MF-01/02 through the burn |
The baseline is computed for your propellant pair and chamber conditions, and its assumptions are printed beside the result — the efficiency is checkable, not asserted.
| IGNITION DELAY | 0.52 s | from the valve command |
|---|---|---|
| Pc PEAK / STEADY | 2.32 / 1.84 MPa | overshoot +26% |
| FLAG RAISED | t+0.42 s | PC-01, automatically |
| CHILL-DOWN | ∇T before valve opening | precursor checked ahead of t+0 |
Start behaviour is classified from the chamber-pressure trace against fixed thresholds, and the gradients in the oxidizer line before the command are read as the precursor they are.
| THRUST VARIATION | CV 3.1% | CV = σ_F / F̄ |
|---|---|---|
| DOMINANT PEAK | 38 Hz | 4.1% of mean thrust |
| PF-02 COHERENCE | 0.71 | at the same peak — feed-coupled |
| RESOLVED BAND | ≤ 200 Hz | bounded by your acquisition rate |
A peak that appears in both thrust and feed pressure is feed-coupled rather than combustion-driven. Above the resolved band the report states no coverage instead of implying it.
| CHANNELS RECORDED | 18 | present through the burn window |
|---|---|---|
| CALIBRATION | confirmed | THRUST-01, PC-01, PF-02, MF-01/02 |
| UNAVAILABLE PARAMETERS | none | every input present for this run |
| ABOVE 200 Hz | no coverage | stated, not implied |
Where a channel is missing samples or a calibration cannot be confirmed, every parameter depending on it is marked unavailable — not estimated, and not averaged over the gap.
| RUNS IN SERIES | 10 | same computation core throughout |
|---|---|---|
| CHARACTERISTIC VELOCITY | 1454 → 1471 m/s | +1.2% |
| THRUST COEFFICIENT | 1.40 → 1.42 | +1.4% |
| SPECIFIC IMPULSE | 208 → 213 s | +2.4% |
| MIXTURE RATIO | 1.41 → 1.35 | −4.3% |
| THRUST VARIATION | 3.6% → 3.1% | −0.5 pt |
Differences are computed over matched steady-state windows. Where a window differs in length, the comparison says so rather than quietly comparing unequal things.
Instead of digging through reports, ask a question. The analyst reads the run's own telemetry, computes the answer, and shows the chart behind it.
Hanabi 1.1 runs today, on real campaigns. Hanabi 2.0 is the next generation — new architecture, built for programmes and teams rather than a single stand.
One workspace per organisation on the new architecture, with stand data processed automatically as it arrives — a firing becomes a finished report without manual steps in between.
Every figure in the workspace opens down to its source data in one click, and the analyst reads your own programme documentation — your engines, your configurations, your previous runs.
Parametric what-if runs and performance prediction across operating points: a proposed change weighed against your own test data before it is committed to a firing, with higher-fidelity flow simulation after that.
From your stand's raw data to a complete analytics package.