Phillip Kingston

Kingston develops frameworks that generate workflows automatically in complex Business Process Outsourcing (BPO) settings, with the aim of cutting cost and raising quality. A workflow, in this framing, is a sequence of tasks that turns a given input into the required output while respecting operational constraints.

He represents workflows as Directed Acyclic Graphs (DAGs): nodes are tasks, edges are execution order. A task holds executable instructions such as code or tool calls, and can include a human review step where quality matters. Decomposed this way, a complicated process reads as a blueprint.

Workflow creation runs in two phases. First a Large Work Model (LWM), a model specialized on work itself, produces candidate workflows from a high-level intention and whatever domain knowledge it can draw on. Those candidates then become a Workflow Graph, and path-finding selects the best route from input to output. The workflow that survives both phases is feasible and cheap to run.

Kingston gives the goal of a process a formal definition. A Workflow Intention is the alignment of the Input, Process and Output elements that describe what a workflow is for. It is read off “Workflow Signals” embedded in business artefacts: the clues in a document or a dataset that point to which input-process-output triple is wanted.

The framework is mathematical. A Workflow Signal is a vector, a Workflow Intention a tensor. Writing the intention as a tensor puts a workflow's components and the relationships between them into a form a machine can read, and one a model can be trained to reproduce.

The implementation is a trainable attention-based system that reads business artefacts and infers the intention behind them. Its modules are separable, one per data modality, so the architecture scales. Structured sources such as a Work Knowledge Graph (WKG) feed domain-specific procedural knowledge into the model, which grounds generation in established business rules as well as in the signals read from the data.

Encoder-decoder transformers pull information out of structured and unstructured sources alike: text, images, scanned documents. Text is tokenized and passed through attention to capture semantics; images go through OCR and convolutional encoders for visual features. Both land in a common high-dimensional space, which is what lets a single system take an input that mixes the two.

He has built an attention-driven generative system that reads forms, documents and images and infers the workflow intention behind them. Modality-specific encoders handle each input type; intra-modality and fusion attention then combine what they produce. An Intention Decoder turns those features into the Input–Process–Output triple that best explains the artefacts. Trained end to end, it takes a stack of business documents and returns the process they imply.

Vision transformers (ViTs) carry the visual side. A ViT treats image patches as tokens and applies self-attention, picking up local and global features without convolution, and Kingston uses them on image-based artefacts such as scanned documents. Models that fuse text and vision embeddings let the system reason across both at once. This is the machinery that turns a heterogeneous pile of business artefacts into something the generator can work from.

Knowledge graphs are how domain and procedural knowledge get into generation. Kingston helps build a Work Knowledge Graph encoding known workflows, their tasks and the relationships between them, gathered across industries. It records which tasks connect to which and which processes are valid, putting business rules and accumulated practice into a structured form.

Reading the WKG during generation is a job for Graph Attention Networks (GATs). Trained over the graph, a GAT suggests which further tasks belong in a workflow that is only partly built. It works by letting tasks attend to their neighbours' features, which surfaces the procedural steps that fit the current context. The generated workflow is then shaped by processes already known to work.

This is retrieval-augmented generation: query the knowledge graph first, then generate against what comes back. The system compares the encoded intention to WKG nodes by embedding similarity and keeps the closest ones. Those tasks and sub-workflows go to the LWM as extra context, so what it writes stays consistent with domain knowledge. Retrieval buys both accuracy and compliance with how the work is actually done.

Picking one workflow out of many needs a cost model. Every candidate path through the workflow graph is scored by a cost function over compute, processing time and any model or tool usage it incurs, combined as a weighted linear sum. With resource and time constraints priced, workflows can be compared on efficiency and not merely on whether they run.

Choosing the execution path is classical graph search, with learning-based methods held in reserve. The implementation runs a modified Dijkstra from the input node to the output node, adapted to charge costs to nodes (the tasks) rather than only to edges. Reinforcement learning is treated as the option for complex or shifting cases; the action space in workflows is large enough that the current system leans on graph algorithms instead, leaving RL for agents that must learn task sequencing in richer environments.

Generated workflows are measured against real-world standards. Kingston has helped build industry benchmarks, among them a test of the Opus system on a medical coding workflow where the specialized models beat general Large Language Models (LLMs) by 29–38% on workflow generation accuracy. Accuracy alone is not the standard he argues for: a workflow is also judged on execution time, resource consumption and adherence to business rules. Judged on all four, it can be trusted in production and not only on paper.

The optimization work led to a quantitative account of how good a workflow actually is. The Opus Workflow Reward treats expected performance as a probabilistic quantity covering the chance of success, the resources consumed and the quality of the output at once. Correctness, reliability and cost end up in one model, which puts competing workflows on a common scale instead of merely sorting the feasible from the infeasible.

Reward has a counterpart: Opus Workflow Normative Penalties, measurable functions over the structural and informational quality of a workflow along four dimensions, Cohesion, Coupling, Observability and Information Hygiene. They catch what raw success rates miss, and they favour workflows that are well-factored, loosely coupled, easy to monitor and disciplined about information.

Reward and penalties together support joint optimization across the trade-off between them, which is enough to assess, rank and select workflows under several quality constraints at once. Because the whole evaluation is quantitative, it can double as a reward signal: a reinforcement learning system can discover and refine workflows against it, closing the loop from generation through optimization to evaluation.

The same optimization-under-constraint thinking carries from digital workflows to physical ones. Kingston co-develops an AI-controlled life support system for firefighting suits, fusing first-principles physical chemistry with neural control. Its physical layer models a semi-closed-circuit breathing apparatus, with carbon dioxide scrubbing by soda lime, humidity control by silica gel and oxygen replenishment, alongside explicit oxygen-safety margins and toxicity thresholds. The result is an 18-state nonlinear model, still tractable enough to run on real equipment.

Above that model sits a layered control architecture: environmental and internal-atmosphere sensing with triple-redundant oxygen measurement, model-predictive control for dynamic resource allocation, and reinforcement learning to propose policies. Every command passes a safety filter built on control-barrier functions before it is issued. In simulation the stack gains 18–34% endurance over PID baselines while holding tighter physiological and fire-safety margins.

Kingston derives the exact conditions for sustaining a breathable atmosphere in a capacity-limited, venting two-gas system, characterizing both the viable region of operation and the best allocation of the two gases under finite storage and venting. Where a conservative heuristic would leave margin unaccounted for, the analysis pins the feasibility boundary down exactly, along with the allocation that goes with it. The guarantees are formal: when life support holds, and how limited reserves should be spent for the longest safe endurance.

When can a supervisory controller be reshaped without changing what it can physically do? Using equality-support covers, Kingston establishes the conditions under which the exact set of reachable branches, meaning the admissible control paths, survives a fixed actuator interface. The results prove that safety-preserving reformulations of a policy leave its true operational envelope intact, which is what trustworthy autonomy in life-critical settings rests on.

In the decentralized setting, several local supervisors each see only part of the system and must jointly decide whether to permit an event. Kingston formalizes control where the authority to veto an event varies with the hidden execution history, and where that authority may itself be concealed from the supervisors. The model covers delegated and revocable control architectures, separating what a command does from whether its issuer was permitted to issue it, and gives precise language for the state-dependent authority constraints found in real actuator and safety-filter interfaces.

A specification is realizable exactly when it is controllable and “legitimately co-observable”: for each forbidden action there must be one supervisor that both knows the action is required and is authorized to veto it across an entire information cell. Kingston proves this, and shows by minimal obstruction that decision knowledge and authority knowledge held by different supervisors do not compose. Beyond exact realization, he extends relative co-observability to construct the supremal safe sublanguage under hidden, history-dependent authority. That yields an exact finite synthesis procedure, and a formal answer to which behaviours a controller can attain at all.