Selected engineering work
Process intelligence from video
Verified active years: 2026
Business-process recordings become documentation, BPMN diagrams, test cases and a searchable knowledge base — semantic embedding with pgvector, hybrid vector + governed SQL retrieval, async job orchestration.
The problem
Business-process recordings need to become useful outputs and searchable knowledge. The backend had to support long-running processing, deliver generated artifacts, and make the resulting information available for follow-up questions.
My contribution
Core AI and backend engineer
Backend and AI integration for test-case delivery; asynchronous job orchestration, repository APIs, auto-ingest, semantic chunking and embedding, hybrid retrieval, task-scoped conversation history and production hardening.
Team / platform context
The video-to-artifacts platform was team-built. My work focused on backend, orchestration and the knowledge repository, not sole ownership of the frontend or the full platform. Aurevia is a separate illustrative portfolio demo.
Try it on this page
Interactive demo
Aurevia is an illustrative browser-only demo using fictional invoice content. Recordings are stored locally, not analysed. Answers use keyword matching, not live AI, vector retrieval or SQL. Demo results are not project outcomes.
Explore the fictional Aurevia workspace
Explore the platform pattern. The separate demo uses fictional content and no live AI.
01 Record a process
The backend stores a recording and creates a tracked job. Upload requests do not wait for AI processing.
02 Generate structured outputs
Workers produce documentation, chapters, guides, diagrams and test cases. Job status and heartbeats keep long-running work visible.
03 Build reusable knowledge
A separate ingest job chunks the generated documentation, creates embeddings and stores searchable knowledge.
04 Retrieve and follow up
The repository combines semantic search with governed SQL and synthesises evidence-backed answers. Conversation history stays scoped to the task.
Team / platform flow above. My contribution focuses on backend integration, orchestration and the knowledge repository.
Explore the fictional workspace →From a recording to reusable knowledge
The team-built platform turns business-process recordings into documentation, chapters, HTML guides, BPMN diagrams, test cases and searchable knowledge. My work focused on the backend, orchestration and repository layer.
Deliver generated outputs
I contributed backend and AI integration for test-case generation and the APIs used to download generated artifacts. Uploads became asynchronous jobs claimed by workers, with statuses and heartbeats providing visibility into processing.
Make outputs retrievable
My knowledge-repository work covered auto-ingest of generated documentation, semantic chunking and embeddings with pgvector. Hybrid retrieval combined semantic search with governed SQL, followed by evidence merging and answer synthesis.
Define the contribution boundary
I contributed repository APIs, task-scoped conversation history and production hardening. The broader video-to-artifacts platform was a team effort; my role was not sole ownership of the frontend or every generated output.
Engineering decisions
Represent processing as tracked jobs
The platform uses asynchronous jobs with statuses, claims and heartbeats in a database-driven control plane. This separates upload handling from longer-running processing and makes worker progress visible without a separate queueing platform.
Combine semantic retrieval with governed SQL
My repository work combined semantic chunking and pgvector embeddings with governed SQL retrieval. The approach supports both text-based knowledge retrieval and structured queries, with evidence merging and answer synthesis.
Keep conversation context scoped to the task
Task-scoped history and repository APIs organise follow-up questions around the relevant work. SQL parameterisation, ingest guardrails and failure metrics support that boundary in the implementation.
Outcomes & evidence
- Contributed backend and AI integration for test-case generation and delivery, including download APIs.
- Built knowledge-repository capabilities covering auto-ingest, semantic chunking, embeddings, hybrid retrieval and task-scoped conversation history.
- Contributed production hardening through SQL parameterisation, ingest guardrails, failure metrics and runtime decoupling.
The evidence here is the documented implementation scope. No measured time saving or retrieval-quality benchmark is published. Aurevia is a separate fictional illustration with no live AI processing.
Engineering takeaway
Generating an artifact is only part of the workflow. Job visibility, delivery APIs and a searchable repository make the output usable beyond the original recording.