project · 2026 · Data Scientist · SAP
Agentic Workflow Platform for Enterprise Operations
A production platform for building and governing intelligent multi-step AI workflows, combining LLM reasoning, structured knowledge retrieval, and external service calls in a single deterministic execution engine with role-scoped governance.
Enterprise operations at scale share one structural problem: the knowledge to handle a complex request exists, but it is scattered across documentation, transaction history, search results, and LLM reasoning. The bet behind this platform is that you can encode expert decision logic as a reusable workflow once, and scale it to every engineer who faces the same class of problem.
High-level architecture
┌─────────────────────────────────────────────────────┐│ Workflow Trigger ││ (case input / user request) │└───────────────────────┬─────────────────────────────┘ │ ▼┌─────────────────────────────────────────────────────┐│ Execution Engine ││ ││ ┌────────────┐ ┌─────────────┐ ┌─────────────┐ ││ │ Record │ │ Knowledge │ │ Reasoning │ ││ │ Fetch │─▶│ Retrieval │─▶│ (LLM) │ ││ └────────────┘ └─────────────┘ └──────┬──────┘ ││ │ ││ Shared Execution Context ◀───┘ ││ (accumulates outputs across steps) │└───────────────────────┬─────────────────────────────┘ │ ┌─────────────┴─────────────┐ ▼ ▼┌──────────────────┐ ┌─────────────────────┐│ Retrieval Layer │ │ Model Layer ││ │ │ ││ Vector Search │ │ In-house Models ││ Knowledge Graph │ │ External Models │└──────────────────┘ └─────────────────────┘ │ ▼┌─────────────────────────────────────────────────────┐│ Registry / Governance ││ Prompts · Agents · Workflows · Policies │└─────────────────────────────────────────────────────┘A typed execution graph
At the core is a workflow execution engine that runs an ordered sequence of typed steps through a shared execution context. Three step kinds carve up the problem space cleanly:
- Record Fetch pulls structured data from systems of record (structured records, logs, entity data)
- Knowledge Retrieval calls external search services, vector stores, or knowledge bases
- Reasoning prompts the language model with the accumulated context to reason, summarize, classify, or generate
Every step reads from and writes into a shared context object that grows as the workflow progresses. This keeps LLM steps grounded: the model always reasons over retrieved evidence, not just the original query.
Avoiding redundant work without sacrificing correctness
Multi-step workflows over live operational data are expensive to re-run from scratch. The platform uses a session-scoped identity scheme that ties each run to its prior results, so the engine can skip steps whose inputs have not changed. The tricky part is consistency: the moment any step in the chain produces a fresh result, every step that follows must also re-run. A single staleness signal propagates forward and never resets, keeping the output coherent without needing a full replay. The practical payoff is fewer LLM calls per session and noticeably lower latency when a user revisits the same session.
Governance as a first-class concern
Access is governed at the use-case level, not globally. Each team operates within its own permission boundary, so contributors in one workflow domain have no implicit access to another. This isolation matters in practice: draft workflows stay contained within the team that owns them until they pass review and are explicitly published.
The registry layer (prompts, agents, tools, workflows, and policies as versioned, reusable artifacts) is what makes governance tractable at scale. The platform ships a centralized prompt repository and a compliance review layer so that prompt changes go through the same review gate as workflow publishing.
Extending the platform
Beyond the core engine, three areas of work expanded what the platform could do.
Onboarding as a skill. New use cases and workflow configurations were previously a manual handoff. Structured onboarding flows were built as reusable skills for an AI coding assistant, reducing the setup friction for teams adopting the platform and making the process repeatable without engineering involvement each time.
MCP interface for the execution engine. Exposing the execution engine via a Model Context Protocol server means external agents and tools can trigger and interact with workflows through a standard interface, rather than bespoke API integrations per consumer.
Visual workflow translation. Existing automation logic spread across n8n graphs needed to move into the platform’s execution model. I worked on translating those visual workflows into the typed step format the engine understands, preserving the intent while gaining the platform’s caching, governance, and context-accumulation properties.
The execution engine connects to a retrieval backend spanning both vector search and a knowledge graph, alongside an LLM proxy layer that routes across in-house and external models. Workflow discovery and agent extensibility are exposed via standard MCP and A2A server interfaces, making the platform composable with the broader enterprise AI ecosystem without bespoke integrations.
Scale and intended reach
The platform is designed to operate at enterprise scale, with the intended audience spanning tens of thousands of engineers and support professionals across a global organization. A single workflow, once authored and approved, becomes available to the entire user base instantly, multiplying the impact of domain expertise without requiring individual retraining or manual knowledge transfer. At that scale, even modest improvements in resolution speed and consistency translate into meaningful reductions in operational overhead and measurable gains in the quality of outcomes delivered to end users.