AI systems engineering

We put models where
being wrong costs money

Most AI work stops at the demo. Ours runs unattended in a live decision path, makes consequential calls in bounded time, and has to explain every one of them afterwards.

We build the containment architecture that makes that safe: the layering, the calibration, the fallbacks, and the audit record. The model is the easy part.

Start a conversation
See the reference system

Built to sit on top of
what you already run

AlphaFlux is our own product: an autonomous system that reads market microstructure, runs a three-architecture model ensemble under a latency budget, scores its own conviction against a historical evidence base, and declines to act when the evidence is thin. It has been live since September 2025.

Everything on this page is drawn from a system we own, operate, and are accountable for. You can read the full architecture, including the parts that were hard.

AlphaFlux system map
Seven layers from tick ingest to order. The ensemble, the calibration layer, the memory architecture, and the per-layer failure design.
Read the architecture

The stack is not about trading

Strip the domain out and the same seven layers describe any system that has to turn a stream of events into a consequential decision, quickly, and defend it later. The vocabulary changes. The contracts between layers do not.

LAYER

ALPHAFLUX

CLINICAL OPERATIONS

CLINICAL OPERATIONS

L1 · INGEST

normalize the event stream

ticks · depth

deterministic replay

vitals · orders · notes

HL7 · FHIR normalization

submissions · third-party data

document extraction

L2 · STATE

maintain the world model

zones with lineage

volume distribution

patient trajectory

unit and staffing state

exposure aggregation

portfolio concentration

L3 · DETECTION

cheap rules before costly inference

setup library

structural preconditions

deterioration triggers

protocol eligibility

appetite and eligibility rules

hard declines

L4 · INFERENCE

ensemble · parallel · weighted join

vision · reasoning · diffusion

conviction, not price

imaging · notes · time series

HL7 · FHIR normalization

documents · history · risk signals

assessment, not price

L5 · EVIDENCE

calibrate · retrieve analogs · abstain

Alpha Echo

calibrated p · EV

comparable cohorts

outcome base rates

comparable risks

loss distributions

L6 · RESOLUTION

deterministic numbers only

levels · position size

deterministic replay

dosing · scheduling

safety constraints

premium · limits · terms

rate tables · guardrails

L7 · ACTION

policy-driven, fully logged

order lifecycle

policy exit

bind · refer · decline

referral to human

submissions · third-party data

document extraction
TWO LAYERS INFER. FIVE COMPUTE. THE BOUNDARY IS THE SAME IN EVERY DOMAIN.
One architecture, three domains. Clinical and underwriting columns are illustrative mappings, not delivered engagements.
The failure mode

Most AI pilots fail as architecture, not as models

The model works in evaluation, then goes into production wired directly between the data and the action. Nothing bounds its latency. Nothing catches it being confidently wrong. Nothing can reconstruct why it did what it did. The pilot gets pulled, and the model takes the blame.
DIRECT
Image
DATA
Image

MODEL

does everything
ACTION
  • latency unbounded
  • no fallback when it stalls
  • confident and wrong is silent
  • numbers hallucinated, not resolved
  • no replay, so no learning
  • blast radius is the whole system
LAYERED
STATE · deterministic, replayable
RULES · cheap filter, hard constraints
INFERENCE · budgeted, ensembled, droppable
EVIDENCE · calibrated, can abstain
RESOLUTION · every number from code
ACTION · policy-driven, fully logged
  • every layer degrades to a safe state
  • a bad model costs a decision, not the system
THE DIFFERENCE IS NOT MODEL QUALITY. IT IS WHERE THE MODEL IS ALLOWED TO SIT.
The same model, wired two ways. Only one of them survives contact with production.

Six things we are unusually good at

Decision path architecture

Deciding which layers may be probabilistic and containing them there

Ensemble design

Matching architecture to sub-problem, parallel fan-out, weighted joins, dissent as signal

Calibration and abstention

Making confidence outputs honest, and building systems that can return nothing

Latency budgeting

Bounded worst case, droppable components, degradation that is always more conservative

Memory architecture

Tiered consolidation, selective retention, retrieval that improves the next decision

Audit and replay

Deterministic reconstruction of any decision, including the ones that produced no action
We also build the ordinary parts well: ingestion, data modeling, evaluation harnesses, and the interfaces the people accountable for the decisions actually use.

Where this architecture pays for itself

Not every AI problem needs it. A summarizer does not. The cost is justified when a wrong decision is expensive and someone will eventually ask you to prove why it was made.
MUST BE EXPLAINED LATER →
THIS ARCHITECTURE PAYS

clinical decision support

grid & process control

underwriting · claims

fraud · AML

contract & compliance review

internal search

content drafting

logistics dispatch

low
high
CONSEQUENCE OF A WRONG DECISION →
IF A REGULATOR, A CLINICIAN, OR A BOARD WILL ASK WHY, THE AUDIT LAYER IS NOT OPTIONAL.
Placement is illustrative. The upper right is where unexplainable AI stops being viable.

What the layering buys, by domain

Clinical operations

The decision
Deciding which layers may be probabilistic and containing them there
What the layering buys
A model that abstains on thin evidence instead of producing a confident number, and a record that shows what was known at the time.

Underwriting and claims

The decision
Bind, refer, or decline, and at what price.
What the layering buys
Pricing that comes from rate logic rather than a model, with model judgment confined to assessment and logged separately.

Industrial and energy operations

The decision
Dispatch, curtail, or hold, under a real-time constraint.
What the layering buys
A bounded latency budget and a fallback that is always more conservative than the component that failed.

Fraud and financial crime

The decision
Block, review, or allow, in the time a transaction takes.
What the layering buys
Calibrated scores that mean what they say, so thresholds can be set on expected cost rather than on intuition.

Contract and compliance review

The decision
Accept, flag, or escalate against policy.
What the layering buys
Hard constraints enforced in code, model judgment on interpretation, and a replayable trail for every flag and every non-flag.

Logistics and field operations

The decision
Route, reassign, or delay against changing conditions.
What the layering buys
Persistent memory, so the system learns which of its own recommendations actually worked.
We also build the ordinary parts well: ingestion, data modeling, evaluation harnesses, and the interfaces the people accountable for the decisions actually use.
Engagement

How we start

Image
01
Decision audit
map the decision pathfind the unbounded parts
Image
02
Decision audit
map the decision pathfind the unbounded parts
Image
03
Build
map the decision pathfind the unbounded parts
04
Handoff
map the decision pathfind the unbounded parts
PHASE 01 IS SHORT, FIXED-SCOPE, AND USEFUL EVEN IF YOU STOP THERE.
We would rather tell you the layering is unnecessary in phase one than sell you a platform you do not need.

The delivery process
is part of the product

AlphaFlux is our own product: an autonomous system that reads market microstructure, runs a three-architecture model ensemble under a latency budget, scores its own conviction against a historical evidence base, and declines to act when the evidence is thin. It has been live since September 2025.

Everything on this page is drawn from a system we own, operate, and are accountable for. You can read the full architecture, including the parts that were hard.

AAOS also runs internally on AlphaFlux, which is how we know it holds up under a system that cannot afford undisciplined change.
Start a conversation
Read the reference architecture