Source material stays attached to the recommendation
The reviewer should be able to inspect which records support a recommendation instead of receiving an untraceable model summary.
Loading portfolio
AI and automation / 2026
SignalWeave
Operational incidents are difficult to investigate when source events, supporting records, recommendations, approvals and model activity are spread across separate tools and cannot be traced back to one accountable timeline.
SignalWeave turns signed operational events into a traceable incident timeline. The current deterministic vertical slice covers tenant isolation, duplicate-safe ingestion, source retrieval, grounded recommendations, approval, audit history and model observability while clearly separating implemented foundations from connector and production-deployment work.
- Role
- Product and systems engineer
- For
- Independent or academic work
- Stage
- Research prototype
Walkthrough
Watch it run
SignalWeave's labelled local demo: one incident traced across GitHub, Zendesk, Jira and Slack, evidence ranked for a question, a rejection blocked until it has a reason, and an approval recorded without any external action.
My part
What I built
- Defined the product scope, domain contracts and repository architecture across API, web, core services, connectors and data.
- Implemented signed and duplicate-safe event ingestion, tenant authorization, source retrieval, grounding checks, approval and audit state.
- Built the responsive Situation Room and source explorer for incident review, record inspection and operator decisions.
- Added security, architecture, release-scope, evaluation and implementation-status documentation so unfinished work remains explicit.
Decisions
The choices that shaped it
Approval is application state
A model can propose a response, but the server owns who may approve it, whether it has already been executed and what audit record is created.
Deterministic behaviour comes before provider integration
The initial vertical slice remains reviewable without OAuth credentials, external connectors or live model inference, which keeps product and security decisions testable.
What has been shown to work
- Domain and API tests cover canonical events, cross-tenant access, webhook signatures, stale delivery and duplicate ingestion.
- Browser checks cover the Situation Room, source explorer and persisted approval workflow across desktop, tablet and mobile layouts.
- Grounding tests reject unsupported facts and the synthetic evaluation baseline remains labelled as synthetic.
Current stage and limits
- The repository is a production-shaped foundation rather than a deployed production service.
- Live OAuth connectors, durable workers, production identity, object storage and model-provider adapters remain external-dependency work.
Screens
From the product
Technical detailOpen the system view
SYSTEM ARCHITECTURE
SignalWeave system view
SignalWeave turns signed operational events into a traceable incident timeline. The current deterministic vertical slice covers tenant isolation, duplicate-safe ingestion, source retrieval, grounded recommendations, approval, audit history and model observability while clearly separating implemented foundations from connector and production-deployment work.Verified sessions or connector contexts establish tenant identity; client-supplied tenant identifiers are not accepted as authority.
Signed ingress normalises external events into a canonical tenant-scoped operational record and rejects stale or duplicate delivery.
Incidents link retrieved source records, recommendations, approvals, audit records and model-execution traces to one reviewable timeline.
Grounding validation checks recommendation claims against the selected source records before approval is available.
External actions remain server-enforced, permission-checked and idempotent.
- 01Input / actorconnects toInterface
- 02Interfaceconnects toCore service
- 03Core serviceconnects toOperational boundary
- 04Operational boundaryconnects toState / outcome
Security and privacy
- Tenant identity comes from verified session or connector context, never ordinary client authority.
- Imported content remains untrusted data and cannot grant permissions.
- Secrets, connector credentials and private operational records remain server-side and are not published.
Built with
More work
Gazelle's KitchenA WhatsApp business gatekeeper that qualifies genuine orders, enforces business and safety rules, contains off-topic or abusive conversations, and only surfaces the small set of cases that genuinely need the owner.PilotSellavie AIAn independent project: multi-tenant conversational-commerce software that connects social messaging with business knowledge, guarded AI, human handoff, orders, billing, analytics and operations.MaintainedAI Automation Control PlaneA runnable multi-tenant control plane for AI-assisted business workflows, combining FastAPI, Next.js, PostgreSQL, Redis, n8n, provider routing, human approval, idempotency, replay, cancellation, audit traces, and deterministic local demonstrations.Maintained