Back to projects
Applied MLGenerative SystemsBackend

ADC Living Identity

The ML layer behind ADC's living AI-driven identity: context embeddings, similarity search, pattern fusion, and preference feedback.

Role
Backend and ML implementation
Focus
Applied ML / Generative Systems
Stack
Python, Embeddings, all-MiniLM-L6-v2

Project narrative

A brand identity that learns from context

ADC's living identity was created to solve a problem that static brand systems are not built for. Amsterdam Data Collective works in a field that changes quickly, where the material of the company is not only a logo, a color palette, or a campaign style, but the accumulated context of its people, projects, reports, culture, and technical practice. The identity needed to express that body of work without falling into the generic visual language that surrounds a lot of AI companies.

The ambition was to let the brand produce visual expressions from meaning. A project description, report, campaign brief, or internal update becomes the input. The system places that input inside ADC's semantic space, finds related pieces of company context, and uses those relationships to guide pattern generation. The output is therefore not random decoration. It is a visual response shaped by what the input is about and how it relates to ADC's own knowledge base.

My work sat behind that translation layer. I helped implement the backend and ML system that turns text into embeddings, searches nearby contextual data, retrieves mapped pattern candidates, and fuses those neighbors into controlled visual variations. When people at ADC select the output they prefer, that choice becomes part of the future retrieval pool. Over time, the identity becomes more than a generator: it becomes a feedback system shaped by collective use.

What makes the project interesting technically is that the machine-learning layer has to support a design idea rather than overpower it. The model is not there to produce infinite novelty for its own sake. It gives the identity memory, context, and evolution while preserving a shared visual language. That is the part I find most compelling: the system treats brand expression as something alive, but still governed by structure.

Technical implementation

How the identity becomes data

01

Contextual training data

ADC source material was distilled into a compact set of contextual data points describing the company's work, culture, people, and values. Those points form the semantic layer the tool searches against.

02

384-dimensional user input

When a user provides context, the text is embedded with the all-MiniLM-L6-v2 sentence-transformer model into a 384-dimensional vector representation.

03

PGVector similarity search

That input vector is compared with the stored ADC context using cosine similarity in PostgreSQL via PGVector, retrieving nearby vectors and their mapped visual patterns.

04

Temperature-scaled pattern fusion

Neighboring patterns are blended into three variations using different temperatures, controlling how strongly the closest or more distant matches shape the output.

05

Preference feedback loop

When a user chooses a favorite output, that selection becomes eligible for future retrieval. Over time, collective preference changes the neighbor pool and steers the identity.

Engineering

Key decisions

01

Connecting abstract brand context to concrete visual output without making the system feel random

Built the semantic retrieval layer around sentence-transformer embeddings and PGVector-backed cosine similarity

02

Distilling ADC's work, people, and values into a compact searchable knowledge base

Used temperature-scaled interpolation to blend neighboring patterns into three distinct variations

03

Turning user preference into a useful feedback signal for future generations

Fed selected outputs back into the eligible neighbor pool so the identity could evolve through usage

Outcome

What changed

  1. 01Supported a living brand identity system that generates context-aware visual expressions
  2. 02Connected static brand assets to a backend that learns from use and preference
  3. 03Helped make the technical ML layer legible enough to support a public design case study

Stack

Tools and systems

PythonEmbeddingsall-MiniLM-L6-v2Sentence TransformersPGVectorPostgreSQLSimilarity SearchCosine SimilarityTemperature ScalingBackend APIs

I build delightful systems.

© 2026 Kristian Ernst.