GlassBox Engine — an agentic ML platform that turns business intent into aligned, production-ready recommender systems, with a human-readable Reasoning Trace on every prediction.
v0.1 · ALPHABUILT ON GOOGLE ADK + GEMINIGOTHENBURG, SWEDEN · 2026
01 / 12
GLASSBOX ENGINE
"
The more powerful ML gets, the less anyone understands it — including the engineers shipping it.
— THE BLACK-BOX PARADOX, OBSERVED AT EVERY SERIES A
02 / 12
GLASSBOX ENGINE
01 · The problem
One paradox. Three failures.
The industry treats the black-box problem as a single issue. It's three — and most teams shipping personalization fail at all of them.
GAP 01
The Alignment Gap
// businesses vs. raw algorithms
Businesses want recommendations that balance revenue, discovery, trust and margin. Off-the-shelf models optimize a single metric — usually clicks — and quietly trade away long-term brand value.
model: maximize(clicks)
cost:trust · margin · LTV
GAP 02
The Knowledge Gap
// the copy-paste trap
New ML engineers ship LLM-generated training scripts they don't fully understand. Notebooks run, metrics look fine — until production drifts and no one can debug which loss function failed where.
notebook: 124 cells
understood:3
GAP 03
The Trust Gap
// users vs. recommendations
Users have learned what an opaque "recommended for you" feels like. Without a why, every suggestion reads as manipulation. They click less, churn faster, and trust transparent competitors instead.
user_42: "why am I seeing this?"
app:…→ traced
03 / 12
GLASSBOX ENGINE
02 · The solution
An agentic ML platform for the team without an ML team.
GlassBox Engine combines ML education, business-goal alignment, simulation, and explainable production recommendations in one workflow — orchestrated by a multi-agent mesh.
Google Agent Development KitGemini 3.1 ProtRPC + PostgreSQL · pgvectorReasoning Trace, end-to-end
01 / DEFINE
Intent Alignment Sliders
▸ Architect Agent
Translate business goals — Profit, Discovery, Trust, Diversity — into a transparent reward function developers can inspect and adjust.
02 / TEACH
The Socratic Mentor
▸ Mentor Agent
An agent-integrated IDE that asks targeted questions about loss functions, layers, metrics and data assumptions — instead of autocompleting.
03 / SIMULATE
The Persona Lab
▸ Coordinator Agent
Synthetic personas and behavior scenarios stress-test recommendation strategies under preference shifts and data drift — pre-deployment.
04 / SERVE
The GlassBox Plugin
▸ Reasoner Agent
Pluggable tRPC API on PostgreSQL + pgvector. Every recommendation ships with a human-readable trace of the signals that produced it.
04 / 12
GLASSBOX ENGINE
PRIMITIVE 01 / DEFINE
▸ Architect Agent
Drag goals. Get a reward function.
Business users move sliders for Profit, Discovery, Trust, and Diversity. The Architect Agent compiles those choices into a concrete, inspectable reward function — and quantifies the trade-off.
Visual interface for non-ML stakeholders to set priorities directly
Architect Agent emits a transparent R(s, a) developers can read, edit and version
Every weight change carries a Trace explaining what it costs and what it buys
ProfitDiscovery
ConversionLoyalty
SafeDiverse
Reward function · R(s, a)
R = α·ctr + β·diversity − γ·return_rate
05 / 12
GLASSBOX ENGINE
PRIMITIVE 02 / TEACH
▸ Mentor Agent · IDE
The IDE that quizzes you back.
Most coding agents autocomplete. The Mentor monitors the code you're writing and asks why — about loss functions, layers, evaluation metrics and the data assumptions hiding underneath.
Catches sampling bias, wrong learning rates, and silent data leaks before deploy
Forces explicit reasoning — closes the Knowledge Gap one PR at a time
Hands-on Socratic teaching, integrated where engineers actually work
21from sklearn import train_test_split
22X_tr, X_te = train_test_split(X, 0.2)
23# ↑ no stratify — but 90% of buyers are US
24model.fit(X_tr, y_tr, lr=0.1)
▸ mentor: Before you run this — what's the prior for stratification when 90% of your label distribution is one country? Walk me through what stratify=y would do here.
06 / 12
GLASSBOX ENGINE
PRIMITIVE 03 / SIMULATE
▸ Coordinator Agent
Stress-test on synthetic users first.
The Coordinator Agent spawns thousands of behaviorally-coherent synthetic personas — and runs preference shifts, data drift, and adversarial scenarios against your model before a real customer ever sees it.
Cold-start without 10,000 real users — design partners ship on day one
Catch fairness and bias edge-cases offline, with full reproducibility
persona_07 · Budget students · 412
persona_03 · Loyal returners · 287
persona_12 · Discovery seekers · 198
07 / 12
GLASSBOX ENGINE
PRIMITIVE 04 / SERVE
▸ Reasoner · tRPC · pgvector
One API call. Recommendation + reason.
A pluggable tRPC API backed by PostgreSQL with pgvector. Every recommendation ships with a Reasoning Trace — auditable, human-readable, and ready to show your user, your PM, or your auditor.
Drop-in TypeScript and Python SDKs — no ML team required
pgvector embeddings + business constraints, queried in one call
Trace cites user signals, product attributes and active constraints
// POST /recommend
{
item: "jacket_2891",
score: 0.84,
trace: [
"denim_affinity: 0.71",
"within_budget_band: true",
"weekly_boost: local_artists"
]
}
Why this → user_42 likes denim · qualifies for free shipping · Discovery weight raised this week to surface local artists.
08 / 12
GLASSBOX ENGINE
03 · Architecture
Specialized agents. One ML lifecycle.
Built natively on the Google Agent Development Kit and Gemini 3.1 Pro. ADK is a strong fit because it's purpose-built for building, debugging, deploying and orchestrating multi-agent systems — exactly the lifecycle GlassBox owns.
▸ Orchestrator
Agent Mesh
Google ADK · Gemini 3.1 Pro
define
Architect
teach
Mentor
simulate
Coordinator
explain
Reasoner
serve
Plugin
01
Native ADK orchestration
Agents are first-class ADK citizens. State, tool-calls and handoffs replay through the Mesh — every decision is auditable post-hoc.
02
One brain per phase
Architect drafts the reward function. Mentor teaches. Coordinator simulates. Reasoner attaches the trace. Plugin serves it. Each agent does one job, well.
03
Reasoning Traces, end-to-end
Every output — a reward weight, a code suggestion, a recommendation — carries a Trace back through the agents that produced it. Nothing is "because the model said so."
09 / 12
GLASSBOX ENGINE
04 · The signature moment
Slide a goal. Watch the loss function rewrite itself.
Architect translates intent into math. Coordinator simulates impact across thousands of personas. You see the trade-off before you ship — not after a quarterly review.
~/glassbox · alignment.live
▸intent.set(profit_to_discovery: 0.62)
architectrecompiling reward function…
Reward function · R(s, a)
R = α·ctr + β·diversity − γ·return_rate
α = 0.38 ↓ from 0.71
β = 0.54 ↑ from 0.22
γ = 0.08
▸persona_lab.simulate(n=1000)
coordinatorspawned 1,000 personas — projecting ↗
— projected impact
Click-through rate
−5.2%short term
Catalog coverage
+38%new designers surfaced
90-day LTV
+11.4%long term
→ architect: Moving toward Discovery surfaces newer designers but drops clicks 5% short term. The Trace says it's worth it. Proceed?
10 / 12
GLASSBOX ENGINE
05 · Why now
Three forces converging — right now.
01
Multi-agent infra is here
Google's ADK ships agent orchestration as a primitive.
The plumbing for debuggable, replayable, deployable multi-agent systems exists today. We're the first ML platform built natively on it — not retrofitted.
02
Gemini 3.1 Pro
Reasoning is finally good enough to teach, not just answer.
Frontier models can hold a Socratic conversation about loss functions, spot bias in a notebook, and translate sliders into honest math. They couldn't last year.
03
Regulation + user revolt
Explainability stopped being optional.
EU AI Act, US state laws, and a generation of users who've learned what manipulation feels like. "The model said so" is now a liability — both legal and commercial.
11 / 12
06 · The ask
Stop guessing. Start aligning.
Alpha is invitation-only — first 200 design partners. We're looking for marketplaces, D2C and vertical SaaS teams shipping personalization without an ML org.
Design partners
20 spots open
3-month engagement. Free during alpha. Co-build the SDK, shape the roadmap, get white-glove onboarding.
Investors
Seed round, Q3 2026
Raising to extend the Mesh — Reasoner v2, eval suite, multi-tenant Persona Lab, SOC 2.
Hiring
Founding eng × 2
Agent infra (ADK / Python) and developer experience (TypeScript / tRPC). Global, hybrid.