
Averroes.ai — Senior Product Designer
Averroes.ai.
End-to-end product design for a B2B AI SaaS platform
Senior Product Designer for an AI inspection platform used in semiconductor manufacturing — defect review, active learning, confidence scoring and annotation. The work included a scalable Figma design system that accelerated design-to-development handoff by 30%, supported by usability testing and heuristic evaluation.
- Duration
- Jun 2025 — Aug 2026
- Tools
- Figma
- Team
- Averroes.ai Product Team
- Year
- 2026
📋Overview
Averroes.ai is a B2B AI SaaS platform for semiconductor manufacturing — computer-vision models inspect wafers and components, and the platform is where engineers review what the model found, correct it, and feed those corrections back into training.
As Senior Product Designer I worked end-to-end across defect-review dashboards, active-learning flows, confidence-scoring interfaces and annotation tools, plus the scalable Figma design system the product is built from.
👩💻Who it is for
Process and quality engineers on the factory floor — domain experts in semiconductor manufacturing, not in machine learning. They are the ones who decide whether a flagged defect is real, and their time is the scarcest resource in the loop.
⚡️The Challenge
The hard problem was translation. A model emits probabilities, confidence distributions and retraining signals; an engineer needs to decide, in seconds, whether to scrap a part, flag a process, or dismiss a false positive. Presented raw, model output is either ignored or over-trusted — and both failure modes are expensive in a fab.
My task was to turn machine-learning concepts into interfaces factory engineers can act on — without a data-science vocabulary, and without hiding the uncertainty that makes the judgement necessary in the first place.
🔍Research & Evaluation
I ran usability testing and heuristic evaluation against the existing product to find where engineers hesitated, mis-read confidence, or abandoned a review. The pattern was consistent: the interface asked people to interpret the model instead of helping them judge the part in front of them.
💡Key Insights
- Confidence needs a decision, not a numberA probability on its own does not tell an engineer what to do; thresholds and defaults have to carry that meaning.
- Review speed is the productDefect review is a high-volume, repetitive task — every interaction saved compounds across a shift.
- Corrections are training dataThe moment an engineer overrides the model is the most valuable signal in the system, and the interface should treat it that way.
STEP 01
Defect-Review Dashboards
The review surface was rebuilt around triage: what needs a human now, what the model is confident about, and what is waiting. Image, prediction, history and action sit in one view so a decision does not require navigation.
STEP 02
Confidence-Scoring Interfaces
Confidence is expressed as a banded, consistently-placed signal tied to the action it implies, rather than a bare percentage. Engineers can see at a glance where the model is sure, where it is borderline, and where their judgement is actually required.
STEP 03
Annotation Tools
Correction tools built for domain experts rather than labelling specialists: a small, predictable set of actions, keyboard-driven, with every correction captured as feedback to the model instead of a local fix.
STEP 04
Active-Learning Flows
The loop that closes the system: corrections queue for retraining, the engineer can see what their input changed, and model updates arrive with a visible before-and-after rather than silently.
STEP 05
Scalable Design System
I built the product's design system in Figma — tokens, components, variants and documented states — so that engineering could implement from a shared source instead of interpreting one-off screens.
📊Results
- 30% faster design-to-development handoffThe scalable Figma design system gave engineering documented components and states to build from, cutting the interpretation step out of handoff.
- Model output engineers can act onConfidence became a banded signal tied to a decision, rather than a number requiring interpretation.
- A closed active-learning loopCorrections made during review feed retraining, and the result of that feedback is visible in the product.
- Validated through testingUsability testing and heuristic evaluation drove the review flow, the confidence model and the annotation tools.
🌟Translate, do not expose
Showing a domain expert the model's raw output is not transparency — it is handing them someone else's job. The design work is carrying the meaning of that output into the decision they are already qualified to make.
✨A design system is a handoff tool
In a product this dense, the system was not a styling exercise. Documented components and states were what let design and engineering move at the same speed.