We combine academic-level statistics with systems engineering. Models are estimated properly, then built into software that holds up under real load. Ten years of it, across finance, logistics and manufacturing.
max E[U(a)]Which action, taken how often, and what does being wrong cost. No modelling starts before the objective is written down.
D = {(xᵢ, yᵢ)}Messy operational reality becomes features, outcomes and constraints, with the data-generating process stated explicitly.
θ̂ = argmin E[L]Fit, validate out of sample, and report the uncertainty. A number without an interval is not an answer.
f(x) → serviceThe estimator becomes a service with tests, budgets, monitoring and drift alarms, running against live traffic.
Same four steps whether the problem is a sales territory, a factory line or an order book. The mathematics changes; the reduction does not.
Shadows Technologies grew out of one combination: engineers building serious production infrastructure and quantitative analysts mathematically modelling financial markets on top of it. Neither discipline works well alone, so we kept them in the same team.
Quant + eng
one team, statistics and systems together
10+ yrs
average experience per engineer, in production
4 sectors
finance, logistics, manufacturing, accounting
0 handovers
the person who models is the person who ships
New systems designed against explicit numbers: data pipelines, pricing and risk services, routing engines, real-time platforms with the analytics wired in.
Existing code and existing models, whatever the language. We measure the real behaviour, quantify the ceiling, then land the fixes with your team in the loop.
A worked example from an insulation installer. Before any routing software gets written, we estimate the probability that a given address converts, using energy label, roof and build age, subsidy eligibility and proximity to recent installations. The route planner then sends crews where the expected value is, instead of where the map looks tidy.
Each pin is an address scored by the classifier. k-means groups the territory into workable day-clusters, and a 2-opt tour orders the highest expected-value stops inside them. Crews drive less and knock on better doors.
Ranking beats guessing by a measurable margin: the top two score deciles hold over half the closes. Features include energy label, roof and build age, subsidy eligibility and proximity to recent installations.
Anonymised outcomes from engagements in trading, freight and manufacturing platforms. Same pattern every time: model the behaviour, remove the waste, prove the delta.
Measured on the client's own hardware, production-shaped load, same functional test suite passing.
Instrument
Traces, counters and samples from the real path.
Quantify
Rank the cost with statistics, not with opinion.
Rebuild
Targeted change sets, benchmarked one at a time.
Guard
Regression gates so the win survives the next release.
Typical first measurable win: within two weeks
Effect sizes, intervals and out-of-sample validation. No conclusion from a single run.
Kernel flags, allocator, network path, query plan. The bottleneck is rarely where it is blamed.
Backpressure, degradation modes and failure drills before go-live, not after.
We work in your codebase, your review process, your CI. No parallel universe branch.
Twenty-year-old C++ or a sprawling monolith: we read code before proposing rewrites.
Your engineers pair with ours. When we leave, the capability stays.
We audit and build in the language your system is already written in. Where we are not the right team, we say so.
One call with an engineer, not a salesperson. Bring a dataset, a trace or just a complaint from operations, and we will tell you what we would measure first.