The practice · Sydney, Australia
Riddhish
Chanda.
I build measurement instruments for domains that resist measurement — cleaning shifts, market regimes, health-service supply, specialist economics. The through-line is not a technology. It is a refusal to ship a number that cannot say where it came from.
Luna Labs is one person, deliberately.
The practice sits at an unusual intersection: machine learning, quantitative finance and actuarial studies, held simultaneously rather than sequentially. That combination is the reason the work looks the way it does. Actuarial training supplies the habit of treating every quantity as a distribution and every model as something that must survive an audit. Quantitative finance supplies the discipline of out-of-sample honesty — that a result which does not reproduce in execution never existed. Machine learning supplies the machinery. None of the three is sufficient alone, and the failure modes of each are most visible from inside the other two.
Working independently is a design decision, not a circumstance. It means the person who audits the raw sensor data is the person who chooses the architecture, writes the training loop, designs the interface and writes the words on the page — so nothing gets lost in a handoff, and no layer gets to assume a guarantee the layer below it never made. It also means the systems here are built to be operated by a small number of people, which is a constraint that improves them.
The name is the method. Luna — the moon, the fixed silent reference by which one navigates. Ichne, the flagship model, is Greek for a footprint: the trace that work leaves behind when nobody is recording it. Both names describe the same commitment, which is to find the anchor first and reconstruct everything else against it.
Method
Five steps, in this order,
every time.
The order matters more than any individual step. Most system failures are traceable to running step three before step two.
-
01
Find the dock
Every system starts by identifying its fixed reference — the known origin that makes everything downstream checkable. In a trajectory problem it is a physical dock. In a trading system it is the single strategy object shared by research and execution. In a data pipeline it is the schema. Without one, errors have nowhere to be measured against and accumulate silently.
-
02
Measure before modelling
The first artifact is an audit of the data as it actually is, not as the specification claims. Sampling rates, units, missingness, duplicate timestamps, semantic drift between sources. This step routinely rewrites the architecture — and it is far cheaper to be rewritten by an audit than by a model that has been training for a week on a lie.
-
03
Compute what can be computed
Geometry, arithmetic and constraint satisfaction are solved by deterministic machinery. Learned components are reserved for the parts that genuinely require them, and language models are used for language. Mixing the two is the single most common way a modern system becomes confidently wrong.
-
04
Render the uncertainty
Every estimate ships with its error, every low-confidence output is allowed to abstain, and the interface makes the difference visible. A model that says "I do not know" on 8% of cases is worth more than one that guesses on all of them and is never audited.
-
05
Publish what failed
Negative results get the same typeface as positive ones. The record of what did not work is the part that compounds — it is what stops the next system from re-running an experiment that has already been settled.
Capability
What that looks like
in practice.
Modelling
- Self-supervised pretraining where labels are scarce
- Multimodal alignment through lightweight projection heads
- Parameter-efficient fine-tuning of small language spines
- Factor-graph state estimation and sensor fusion
- Calibrated confidence and abstention policies
Quantitative
- Event-driven backtesting with live/research parity
- Conviction-scaled position sizing and drawdown control
- Volatility-regime detection and signal gating
- Portfolio state, execution and broker integration
- Stochastic loss and reserving models
Engineering
- Ingestion pipelines resilient to upstream reshaping
- Entity resolution and geospatial joins at scale
- Queue-and-worker architectures for long jobs
- Scale-to-zero cloud design priced for idle
- Protocol servers and agent tooling
Voice
The four rules
that rule things out.
-
01
Evidence over adjectives
A claim without a measurement behind it is a mood. Numbers carry provenance or they do not ship.
-
02
Uncertainty is rendered, never hidden
Every estimate arrives with its error. A confident wrong answer is the most expensive artifact a system can produce.
-
03
Geometry is computed, language is generated
Models write prose. They do not write coordinates. Every figure a model emits is machine-checked against its computed source before display.
-
04
State negative results plainly
What failed is the more useful half of the record. Findings that contradict the plan get the same typeface as the ones that confirm it.