Enterprise AI adoption is accelerating—but so are the risks. From ethical lapses to irrelevant outputs, traditional LLM pipelines struggle with alignment, especially when static rules or prompt engineering are the only lines of defense. What if your AI could learn to stay on-topic, aligned with enterprise values, and semantically coherent—all while adapting over time?
Apr 29, 2025
Why a New Approach? Enterprises deploying language models often face the same challenge: how to ensure responses stay on topic, coherent, and aligned with business goals—without drowning in prompt engineering or moderation scripts.
Apr 29, 2025
Singular Value Decomposion (SVD) is a well known method for latent semantic analysis. When applied to BERT contextual embeddings SVD produces three components: U, Σ, and V. The eigenvectors in V represent distinct semantic patterns - each one captures a different aspect of meaning in the text. The eigenvalues in Σ tell us how significant each pattern is, effectively showing us what is semantically important and where the semantic “holes” are - the gaps in meaning that separate different semantic clusters. This elegant mathematical decomposition reveals the fundamental building blocks of meaning in text, creating a natural hierarchy of semantic patterns that can be analyzed through linear algebra and externalised a semantic knowledge graph.
Mar 30, 2025
Generative AI has high utility for generating metadata for both structured and unstructured data and is relevant in the storage domain where data discoverability drives the value of data across the enterprise including for downstream AI projects.
Dec 4, 2024
Operating AI across different clouds and execution engines becomes complex and difficult to maintain with cloud native tools as the number of different integrations between systems proliferates at scale. OpenShiftAI provides a cohesive hybrid, multi-cloud AI platform that enables enterprises to separate concerns between pipeline orchestration and workload execution reducing complexity in the data and governance subdomains and enabling enterprises to operate AI at scale.
Nov 25, 2024
The watsonx.ai BaseChatModel supports integration with LangChain for building LangChain Tool-Agents. The following code demonstrates use of the LangChain watsonx BaseChatModel to construct a Tool-Agent. The application logic follows: (1) a call to the language model to determine which tools to invoke; (2) the programmatic invocation of the selected tools (3) a final call to the watsonx language model with the response from the tools.
Jul 13, 2024
Q*, a hybridisation of Q-learning and the pathfinding algorithm A*, has the potential to enhance the inductive bias of a language model in tasks that demand certain types of reasoning. An implementation of Q* is described here https://lnkd.in/giMTvSQR and implemented with a watsonx language model here https://github.com/jamesdhope/q--deliberate-planning-watsonx with the following parameters and adaptions:
Jul 10, 2024
Whilst Monti Carlo Tree Search and Q* are promising approaches for aligning and guiding general purpose language models in a specialised domain, MLOps (or LLMOps) remains essential for maintaining models that are susceptible to drift. This is a particular concern in ecosystems where agents with smaller, specialised models and the environments they are deployed into are continously evolving, as these models are comparatively more susceptible to data drift than larger, general purpose models due to their relatively narrow training distribution. Additionally, in cascading heterogeneous agentic architectures out-of-distribution (OOD) inputs/outputs have the potential to propagate and proliferate from agent to agent.
Jun 19, 2024
A key challenge in language model applications is managing the dependency on language model prompts. Changes to the data pipeline, the model or the data requires prompts to be re-optimised. DSPy is a framework for algorithmically optimizing LM prompts and weights that separates the flow of a language model application from the parameters (LM prompts and weights) of each step and provides LM-driven algorithms that can tune the prompts and/or the weights of your LM calls, given a metric you want to maximize. DSPy introduces signatures (to abstract prompts), modules (to abstract prompting techniques), and optimizers that can tune the prompts (or weights) of modules.
Apr 7, 2024
NeMo-Guardrails is an open-source toolkit that allows developers to add programmable guardrails semantically matched on utterances to LLM-based conversational applications. NeMo-Guardrails can be easily integrated with watsonx.ai models using LangChain’s WatsonxLLM Integration.
Feb 27, 2024