AI · ANALYTICS · ARCHITECTURE

Understand the technology.
Decide on business value.

I connect quantitative methods, language models and emerging architectures with business tasks. The right choice follows the quality you need, your data and a viable operating model.

Technical perspective · 24 October 2026

01 / BUSINESS ANALYTICS

Understand numbers. Support decisions.

ANALYTICAL METHOD

Machine learning & predictive analytics

Transactions, customer interactions and operational data support patterns, segments and forecasts. Methods such as gradient boosting provide important baselines for structured data.

Business connection

Customer segmentation · demand · anomalies

Selection criteria & sources

Value over simple models, data quality and stability over time.

ANALYTICAL METHOD

Forecasting, simulation & causal analysis

Time-series models support planning. Simulations make scenarios comparable. Causal methods examine intervention effects under explicit assumptions. Forecasts, scenarios and causation answer different questions.

Business connection

Treasury · capacity · pricing · marketing impact

Selection criteria & sources

Test periods, uncertainty, control groups and identification assumptions.

02 / LANGUAGE & GENERATION

Process language. Generate content.

Architecture describes how a model is built; autoregression and diffusion describe how outputs are generated. A Transformer can also be used inside a diffusion model or NCM.

MODEL FAMILY & ARCHITECTURE

LLMs & Transformers

Large language models process and generate language and code. Many use Transformers: attention weights relationships within context. Autoregressive language models predict the next token from the preceding context.

Business connection

Research · documents · knowledge access · code

Selection criteria & sources

Task quality, source grounding, context requirements, data handling and cost.

MODEL ARCHITECTURE

Mamba & state space models

Mamba processes sequences through a continuously updated state. An input-dependent selective mechanism controls which information is retained. At fixed model and state sizes, computation scales linearly with sequence length.

Business connection

Long sequences · time-dependent signals · model comparison

Selection criteria & sources

Quality, runtime and memory requirements on the specific task and hardware.

GENERATION METHOD

Diffusion for images and language

Diffusion models learn to reconstruct corrupted data step by step. The principle also applies to language: LLaDA, for example, iteratively recovers masked tokens. Multiple positions can be processed within a single step.

Business connection

Image generation · alternative text generation

Selection criteria & sources

Output quality, number of steps and actual runtime.

Four principles, clearly distinguished.

Autoregressive

Context → next token → next token

Generate sequentially from the preceding context.

Diffusion

Corrupted content → refinement → output

Reconstruct multiple positions over repeated steps.

NCM

Coarse codes → finer codes → text

Generate level by level, in parallel within each level.

JEPA

Observed context → predicted representation

Predict in an abstract feature space.

03 / EMERGING APPROACHES

Assess emerging model approaches.

RESEARCH & EVALUATION

JEPA · Predictive representations

Joint-Embedding Predictive Architectures predict missing or future content in an abstract representation space. I-JEPA and V-JEPA demonstrate this for images and video, without having to generate every detail as pixels.

Business connection

Visual understanding · video · planning physical processes

Selection criteria & sources

Transfer to the use case and available training and evaluation data.

NEW RESEARCH · OCTOBER 2026

NCM · Next-Concept Modeling

NCM represents text as a hierarchy of learned codes. A Transformer generates these from broad structure to fine detail, in parallel within each level. The approach explores efficient text generation with few sequential levels.

Business connection

Research examples: translation · summarisation · text generation

Selection criteria & sources

On XSum, the authors report 31× faster generation than the strongest tested diffusion baseline at comparable ROUGE-L, measured on an NVIDIA H100 SXM with 80 GB. This applies to the tested models and settings; readiness and business value require separate evaluation.

04 / APPLIED AI

Turn models into useful applications.

SYSTEM DESIGN

RAG, chatbots & AI agents

Retrieval-augmented generation connects a language model with relevant company sources. A chatbot makes these functions accessible through dialogue. Agents add tools and executable steps, making the model part of a concrete workflow.

Business connection

Internal knowledge systems · research · assisted workflows

Selection criteria & sources

Relevant evidence, permissions, result verification and approvals for consequential actions.

05 / DATA & DEPLOYMENT

Good models need
a sound foundation.

Big Data adds volume, variety and processing demands. Whether you use structured transactions or unstructured documents, data quality, access, integration and evaluation determine what is useful.

Cloud

Approved services and managed infrastructure, assessed for data processing, integration and ongoing cost.

Private Cloud

Dedicated environments for specific isolation, network and operating requirements.

On-Prem

Local operation where appropriate, with hardware, model quality, maintenance and staffing included in the decision.

My role: define the business task, compare approaches, test assumptions and connect the right implementation expertise.

Discuss your task ↗