Black-box optimization

Stop guessing.
Start optimizing.

OTHERS GUESS. INSPECTOR KNOWS.

A unified workbench for derivative-free optimization. Describe the problem. Let Inspector pick, tune, and race the right algorithm — so you converge in fewer evaluations and ship with confidence.

Get started Try Playground
trusted by teams at labs · hedge funds · hw companies
Real-world use cases

Wherever evaluations are expensive.

Any domain where each function call costs money, compute, or wet-lab time. Inspector ships with templates for the ones that matter.

ML · AutoML
Hyperparameter tuning
Kill the grid search. Converge in 50 trials what would take thousands.
objective: maximize val_auc dim 7–40 · budget 50–500
Life sciences
Drug & molecule discovery
Search chemical space when each evaluation means days of lab work.
objective: maximize binding_kd discrete · d ≈ 100
Hardware
Chip & circuit design
Hit timing and power targets when each synthesis run takes hours.
objective: min power @ freq mixed-int · d 15–60
Aero · CFD
Aerodynamic shape design
Optimize airfoil control points when each CFD sim eats an hour of GPU time.
objective: minimize drag_coef continuous · d 20–80
Problem builder

Describe it. We'll formalize it.

Type the problem as you'd explain it to a colleague. Inspector extracts variables, constraints, and routes it to the right solver.

Tune 7 hyperparameters of a CatBoost model to maximize AUC on a 200k-row dataset, budget 80 trials.
examples → minimize drag coefficient of a wing profile schedule 24 tasks across 4 machines pick 12 of 80 features for a logistic model
Extracted problem · parsed in 0.41s
mixed-integer recommended: SMAC
Objective
maximize auc_roc
Dim / Budget
7 · 80 trials
Noise
stochastic · cv=5
Inferred variables
learning_rate · float · [1e-3, 0.3] log depth · int · [4, 10] l2_leaf_reg · float · [1, 10] bagging_temperature · float · [0, 1] random_strength · float · [0, 10] border_count · int · [32, 255] grow_policy · cat · {Depthwise, Lossguide}
AI advisor

An advisor that actually knows the literature.

Ask which algorithm fits your setting. Get a reasoned recommendation backed by benchmark data across 2,400+ problems.

I have 50 binary variables and a budget of 200 evaluations. What should I use?
advisor
DISCRETE

Encode your 50 variables as a flat 0/1 integer vector — one entry per binary decision. Your objective function receives this vector and returns a single scalar. With a budget of 200 evals, start with 32 random samples to cover the space before optimization begins.

confidence 89%
suitability · 5 axes
recommended
SPEED SCALE SAMPLE EFF. CONSTRAINTS EASE
Python · pip

One import. The right algorithm — picked for you.

Tell Inspector your problem's domain and scale; we pick the best black-box optimiser from 13 curated algorithms. You never need to know which one ran.

quickstart.py $ python quickstart.py
1from inspector import Problem, Space, solve
2 
3space = Space(
4    lr=Space.log(1e-4, 1e-1),
5    depth=Space.int(2, 12),
6    activation=Space.cat(["relu", "gelu", "silu"]),
7)
8 
9p = Problem(train_and_eval, space, direction="maximize", budget=200)
10r = solve(p, hint="mixed-small")
11print(r.best) # {'lr': 0.0031, 'depth': 7, 'activation': 'gelu'}
$pip install inspector
13 algorithms
◆Apache 2.0
2.4k benchmarks