AIAGENTDEV
AIAGENTDEV Workspace in practice · A worked example

Featurespace forges a fraud model only it could build.

Cambridge, UK Financial markets, real-time risk & quant Tier B On-prem

The same move AIAGENTDEV runs for any organisation, shown end to end inside one company's boundary. Point Crucible at Featurespace's own behavioural transaction data, forge it into a small owned model, an Ingot, hold it in Continuum as one shared memory, and deliver it through AIAGENTDEV Workspace. Featurespace builds adaptive behavioural analytics for real-time fraud and risk; here is how the foundry gives it a model that speaks only its own data.

Your data→ Crucible forges→ an Ingot→ held in Continuum→ delivered in AIAGENTDEV Workspace→ on AIAGENTDEV
Video · 16:9 Video · full-bleed 2560 × 1440 px Featurespace's fraud model, forged in Crucible and served through AIAGENTDEV Workspace, the worked example in motion.
The pipeline

Inside Featurespace's boundary

Six steps, run entirely within Featurespace's own walls. Crucible selects the open base, trains it on the issuer's data and pours the result into a small owned model. Nothing proprietary leaves the perimeter, and AIAGENTDEV Workspace plus Continuum put it in front of every risk team at once.

01 · Enterprise data, ingested in-workflow
Your data
Transaction time-series Behavioural profiles Fraud labels
Behavioural transaction data kept inside the issuer's boundary.
→
02 · Open-source base, auto-selected
SmolLM3-3BAnalysing...
A small open model for real-time behavioural scoring.
→
03 · Crucible
Fine-tune & distil
Trains the base on your data inside your boundary, then pours it into a small model.
→
04 · The Ingot
feat-behaviour
A small owned model, built around behavioural transaction profiles and confirmed-fraud labels. It never leaves your walls.
→
05 · Served by
API endpoint
Served inline for real-time decisions.
→
Auto-selection engine
Selected · SmolLM3-3B
Signals it reads
Transaction time-series Behavioural profiles Fraud labels Outcome · real-time scoring Constraint · inline latency, on-prem

The engine reads the data types brought in and the outcome the customer is after, then matches them to the open model that fits the job at the lowest cost to serve.

Candidate base models, scored to fit
SmolLM3-3B 0.94
Ministral 3 8B 0.78
Mistral Small 4 0.61
Qwen3-8B 0.52
Selected SmolLM3-3B: the smallest model that clears the scoring and latency bar, so it runs inline, on-prem, at the point of decision.
Step 01 in depth · Data ingress

Bring your data in. AIAGENTDEV Workspace stages it for the forge.

Drop documents, databases and structured exports straight into the AIAGENTDEV Workspace. The ingress engine reads each file, detects its type and normalises it into a staging set that Crucible trains and refines on, all inside your boundary.

AIAGENTDEV Workspace featurespace / crucible / ingest Ingress engine
Drop files to ingest
Documents, databases, structured tables, semi-structured and raw. Or .
PDF · DOCX CSV · Parquet · XLSX SQL · DB dumps JSON · XML · YAML Audio · Images · Raw
Ingress engineIdle
Nothing ingesting. Drop files, or load a sample set.
Staged for training0 files
Staged files land here, normalised and ready for Crucible.
Staged and ready, inside your boundary. Send to Crucible
Step 05 in depth · Serving

Call feat-behaviour like any model.

The forged Ingot lands in Featurespace's own model library and is served as a secure API endpoint inside the boundary. Every risk system scores against it in real time, and Continuum keeps the confirmed-label memory shared across the desk.

AIAGENTDEV Workspace featurespace / models 5 serving
Model library5 Ingots
POST https://api.aiagentdev.si/v1/ingots/feat-behaviour/score
Runs against the in-boundary endpoint
Response 200 OK review 8 ms feat-behaviour · 2026.08.3
Net-new capability · what they can do now that they could not before

Fraud models adapting continuously from confirmed cases, ahead of a slow retrain cycle.

Outcome

Fraud models that were retrained on a slow cycle now adapt continuously from the issuer's own confirmed cases.

In production

Real-time scoring, on the bank's own model.

The fraud Ingot runs inline at the point of decision, scoring every transaction against Featurespace's own behavioural data, inside its boundary.

4:3 Image · half slide 1280 × 960 px Screenshot: the fraud model scoring live transactions in the AIAGENTDEV Workspace console.

The foundry, step by step

Featurespace brings its own data to Crucible: behavioural transaction data kept inside the issuer's boundary. Crucible automatically selects SmolLM3-3B as the open-source starting point, a small open model suited to real-time behavioural scoring, fine-tunes it on that data entirely inside Featurespace's boundary, and distils the result into feat-behaviour, a small owned Ingot. The Ingot is served through a secure API endpoint, inline for real-time decisions, and Continuum makes it shared org memory: every risk team works from one confirmed-label memory that keeps adapting to emerging fraud.

01
Bring your own data
Behavioural transaction profiles and confirmed-fraud labels stay inside the issuer's perimeter.
02
Crucible picks the base
SmolLM3-3B, chosen for low-latency behavioural scoring at the point of decision.
03
Forge the Ingot
Fine-tune and distil into feat-behaviour, a compact owned model that never leaves the walls.
04
Serve and share
An inline API for real-time decisions, with confirmed labels flowing back into Continuum for every team.
One move, every vertical

See the same foundry across one hundred outcomes.