The operating manual for production agents.
Concepts, primitives, and implementation notes behind the Group e-media information AI stack: data substrate, runtime, evals, and operational feedback loops.
Agent Runtime
The execution engine that turns an AI agent from a chat-window demo into a long-running, event-driven, restartable process you can trust with real operations.
Workflow Evals
The test suite your AI workflows have to pass before any change reaches users — measuring quality, latency, cost, and safety on real production data instead of vibes.
Source Graph
A navigable map of every system your data lives in — schemas, documents, code, tickets, events, owners, and permissions — so an AI agent can find the right source and respect the right access boundary.
Data substrate
6 articlesSource Graph
A navigable map of every system your data lives in — schemas, documents, code, tickets, events, owners, and permissions — so an AI agent can find the right source and respect the right access boundary.
Vector Search
How an AI agent finds the right document, chunk, or row to ground its answer in — and why the part that matters is the pipeline around the database, not the database itself.
LLM-Ready Knowledge Base
A company knowledge base built so an AI system can cite real answers from it — sourced from documents, tickets, code, conversations, and structured records; chunked, embedded, permissioned, evaluated, and kept fresh on AWS.
Lakehouse Architecture
An architecture that combines data-lake economics (cheap object storage, open file formats) with warehouse guarantees (ACID transactions, schema evolution, time travel) so analytics, AI retrieval, and machine learning all read from the same trusted tables.
Ingestion Contracts
Explicit agreements between a data source and the systems that depend on it — what shape, how fresh, who owns it, what counts as broken — so pipeline failures become attributable instead of mysterious.
Data Quality
Contracts, validation, lineage, freshness, and ownership for the data your AI reads from — not a one-time cleanup project, an ongoing operating discipline.
Agent runtime
4 articlesAgent Runtime
The execution engine that turns an AI agent from a chat-window demo into a long-running, event-driven, restartable process you can trust with real operations.
MCP Tool Registry
A governed catalog of every tool an AI agent can call — your APIs, your databases, your internal systems — with typed schemas, permission scopes, audit trails, and the standard protocol (MCP) that turns 'we exposed it to the LLM' into 'we know exactly who called what when'.
Human Approval
Approval gates that put a human in the loop where correctness, risk, or accountability actually require human judgment — designed as part of the workflow, not as a panic button bolted on after launch.
Workflow Runtime
The engine that runs an AI agent workflow as a durable, observable, restartable process instead of a one-shot script — what separates an agent demo from an agent deployment.
Evaluation
4 articlesWorkflow Evals
The test suite your AI workflows have to pass before any change reaches users — measuring quality, latency, cost, and safety on real production data instead of vibes.
Self-Optimizing Agents
AI workflows that propose, score, and promote their own variants — prompts, models, retrieval policies, tool budgets, generated code — under measurable constraints instead of intuition or vendor leaderboards.
Prompt and Model Diffs
Side-by-side measurement of a candidate prompt or model against the current production version on the same eval set — the unit of safe change in a serious AI workflow.
Promotion Gates
The thresholds an AI change must clear before it reaches production — quality, latency, cost, memory, safety — enforced by CI, not by hope.
Operations
3 articlesModel Routing
How an AI system decides which model to call for each step — based on privacy, cost, latency, quality, and what happens when a provider goes down.
Agent Observability
Trace-level visibility into every model call, retrieval, tool invocation, decision, approval, and failure inside an AI workflow — the substrate every other discipline (evals, optimization, governance) reads from.
Governance
The policy layer for what an AI system is allowed to read, call, decide, and ship — encoded as configuration the runtime enforces, not as a document on a shared drive.
Research notes
8 articlesConversation Intelligence
Turning every approved conversation — support, email, team chat, customer messaging, voice, sales — into structured signal you can act on, instead of anecdotes that evaporate when a ticket closes.
Chat Orchestration Runtime
The end-to-end architecture of modern conversational AI systems: model-agnostic, client-agnostic, plugin-driven runtimes that coordinate intent, context, retrieval, tools, reasoning, reflection, memory, and rendering — with the LLM as one interchangeable component, not the system.
AI-Native Dashboards
A study on conversational, adaptive, living dashboard interfaces — workspaces that begin as a blank canvas with a single conversational input and build themselves in real time as the user asks, persisting widgets, layouts, and memory across sessions.
Prompt-Native Widgets
Generative, context-aware dashboard components whose logic and rendering are defined by natural language prompts rather than hardcoded configurations — runtime-generated analytical surfaces that retrieve, reason, link, and adapt instead of merely displaying.
Production Agent Interfaces
The chat surface as an operating console — knowledge bases plugged in, tools connected, agents on a roster, with real-time visibility into context budget, token spend, model choice, and concrete savings opportunities. The interface that lets a team actually run an agent in production, not just demo one.
Conversation Listeners
Opt-in listeners that capture conversations from every channel an organization uses — support, email, team chat, customer messaging, webchat, sales tools, voice — and route them into the signal-extraction pipeline with consent and retention rules attached.
Signal Extraction
Turning raw conversation transcripts into structured fields — intent, subject, sentiment, CSAT, tool performance, product mentions — that downstream systems can query, dashboard, and act on.
Conversation Forensics
Incident detection and root-cause analysis on human↔agent conversations — replaying threads, reading the context around negative sentiment, extracting whether the user actually resolved their problem, and turning the answer into a learning artifact the system can use next time.
Capability map
4 articlesSynthetic Personality
Before an AI agent can be useful to anyone, it has to be something — a coherent identity that holds up across users, sessions, and adversarial pressure. This is the research track that defines what that means and how to keep it stable.
Agent Memory
How an AI agent remembers the user it serves — what they said before, what they prefer, what context not to repeat — without that memory drifting the agent's behavior for everyone else.
Skill Distillation
How an educated AI agent — the one with the codebase, the tools, and the tacit context that lets it succeed — distills its competence into transferable skill documents a virgin agent can run from scratch.
Closed-Loop Knowledge
How an AI system gets durably better at its job — not by being smarter, but by routing every production failure into either a knowledge update, an eval case, a workflow patch, or a documented exception with a named owner.