2,847 features. 12 selected. 0.8ms inference. That's how AIDflow selects your scorecard features.
A black box in a regulated environment isn't a model — it's a fine waiting to happen. AIDflow treats every model as an auditable legal artifact. It automates the pipeline mechanics while keeping a human in control of every decision.
3/4 features promoted to scoring model
region_code → excluded (low predictive lift)
A black box in a regulated environment isn't a model. It's a fine waiting to happen.
Automated ML tools optimize metrics in isolation and hand you a black box. In Fintech, Banking, Insurance — that's a liability, not an asset. Your model risk committee needs to see every coefficient. Your auditor needs to replay every decision. Your CRO needs to defend it to the ECB.
Three Layers. Zero Coupling.
Swap algorithms without touching the API.
The state schema and reporting pipeline remain unchanged.
Worker
Interface LayerUser-facing API. Validation, state management, registration, UI. The layer your team touches.
Model
Adapter LayerParameter resolution. Algorithm switching. Fit options. Switch algorithms by passing one parameter.
Fitter
Engine LayerPure computation. No UI, no context. Data in → model out. This is where the math lives.
# Switch algorithms without touching the API from aidflow import ProjectContext ctx = ProjectContext() ctx.genesis( project = 'DEMO', target = 'is_churn' ) ctx.params(logging=1, verbose=2) w = ctx.workers.agent2_predictor() w.fit(fx_choice='LassoGLM', fx_params=(dict=(alpha=0.01))) # works w.fit(fx_choice='ElasticGLM', fx_params=(dict=(eps=1e-3))) # also works — same API w.fit(fx_choice='LogRegGLM', fx_params=(dict=(tol=1e-4))) # also works — your choice
Genesis → Training → Champion
Six stages. Full state preservation. Human at every decision point. Click to advance.
Data Ingestion
FAIflow transforms raw data into structured, context-rich features. Zero-copy Arrow IPC streaming. Panel data converted via snapshot sampling. Future leakage auto-detected and prevented.
Traditional ML vs AIDflow
The difference isn't accuracy. It's control.
| Dimension | Traditional | AIDflow |
|---|---|---|
| Data flow | Variable passing (df1 → df2 → ... → dfN) | Context registration (semantic lineage) |
| Architecture | Monolithic script | Worker → Model → Fitter (swappable) |
| Extension | Rewrite the function | Define a Fitter, register it. Done. |
| Stability | Manual checks, date slicing | Built-in temporal phases (pre-training) |
| Memory | RAM-bound. Sample or add hardware. | Arrow offload. Runs on any machine. |
| Compute | multiprocessing or manual joblib | Local → Ray cluster (one param) |
| Reproducibility | "Works on my machine" | Context-locked. Run-to-run identical. |
| AI Orchestration | Manual step-by-step | Single-call pipeline → production DAG |
What AIDflow Does Differently
Not more algorithms. More control, more transparency, more proof.
Human-in-the-Loop
Every phase is a decision point. Pause, inspect, re-configure, re-run — without losing state. You direct. The engine executes.
Ensemble + Selection
Four orthogonal agents compete. An objective scoring system promotes the champion. No hand-picking. No confirmation bias.
Audit-Ready
Per-step standardized artifacts: stability indices (PSI/CSI), lift tables, calibration, feature lineage, SHAP vectors.
Self-Contained Scoring
The exported artifact encapsulates all encoding and model state. Inference requires no training data or prior pipeline state.
Memory-Aware Relay
Offloads intermediate state to disk via Arrow. Run pipelines on a 16GB machine. ctx.params(offload=1)
Distributed Compute
Swap from local multiprocessing to Ray cluster for multi-node scaling. pip install aidflow[ray]. No pipeline changes.
A human directs.
An AI agent executes.
The output survives committee review.
AIDflow is also an AI-orchestrable endpoint — an endpoint that lets LLM agents drive the full pipeline (genesis → training → champion) and return a deployable, regulatory-grade scorecard without human intervention.
If your scoring model must survive a committee review, AIDflow is your development environment.
# Single-call pipeline → production scorecard from aidflow import ProjectContext ctx = ProjectContext() ctx.genesis( project = 'DEMO', target = 'is_default' ) ctx.params(logging=1, verbose=2) ctx.flow.process('data') # Ingest & Screen data to AIDflow ctx.flow.process('encoding') # Rigorous transformation to WOE ctx.flow.process('training') # Ensemble: 4 agents compete ctx.flow.process('selection') # Assessment and promotion of champion # Human-in-the-loop: pause at any stage ctx.params(offload=1) # 16GB machine? No problem. data = ctx.data_io.load('book_2024.parquet') score = ctx.predictor.champion_predict(data) # Output: deployable, regulatory-grade scorecard report = ctx.workers.champion_predictor.report display(report.show()) # Full scoreard report print(report.summary()) # Summary → AUC 0.847 · PSI 0.042 · 12 fes
Your next model committee
starts in Q3.
We'll run your scoring data through AIDflow. You get the feature selection log, the binning map, the coefficient table, and the stability report. Free. No obligation. No sales call.