Conrad van Coller

What it does

When a developer asks Claude to write Fenergo code, Claude searches the local index first and grounds every response in authoritative documentation. The corpus covers ~2,900 Developer Hub pages, hundreds of Swagger / OpenAPI specifications, and ~5,000 quality-curated answers from the internal Stack Overflow for Teams instance. Retrieval runs entirely on the developer's laptop, with no external API calls and no data leaving the organisation.

Functional consultants use it the same way through Claude Desktop. When they need a specific piece of information, Claude answers from the grounded corpus and surfaces direct links to the relevant sources for verification and further reading.

Agentic research pipeline

For broad integration requirements, the system runs an internal multi-stage pipeline rather than a single search. A rule-based decomposition step tokenises the requirement against a corpus-derived vocabulary table covering Fenergo entities, API clusters, and subsystems, then emits 4-6 targeted sub-queries. Those queries run in parallel, results are fused via Reciprocal Rank Fusion, and chunks are re-ranked by cosine similarity against the original requirement embedding. Narrow lookups take a fast path; complex requirements trigger a deeper retrieval round when coverage is low.

The pipeline then synthesises a structured implementation plan with cited source URLs using MCP Sampling, which routes the synthesis request through whichever model the developer already has open in Claude Desktop or Code. There is no Anthropic API key in the repo and no second LLM client. The system uses the AI already in the room.

Architecture

A small Node.js MCP wrapper fronts grounded-docs, an extended build of the open-source arabold/docs-mcp-server. The wrapper re-exposes its tools with Fenergo-scoped descriptions, enriches results with canonical source URLs, and emits OpenInference TOOL spans asynchronously so telemetry never blocks retrieval. The whole stack ships as a single Docker Compose project that a developer brings up with one command.

Indexing & distribution

Documentation is chunked along Markdown structure so each chunk holds one self-contained idea: a single endpoint, a concept, a Q&A pair. Chunks are embedded locally with snowflake-arctic-embed2 and persisted into a SQLite vector store. The pre-indexed store ships as a compressed archive in Git LFS, and a scheduled job rebuilds it weekly. Developer machines simply git pull to refresh, with first grounded query under five minutes from a clean clone.

Telemetry & observability

A companion Phoenix OSS deployment captures every MCP tool call, giving engineering leads visibility into which docs are retrieved, retrieval latency, token consumption, and per-user activity. Phoenix is deployed to Azure Container Apps via a small Aspire project and surfaced back into Claude Desktop as an MCP server, so leads can query traces in natural language.

What it does

Fen-X holds client lifecycle and compliance data in a NoSQL backend that is not designed for high-volume analytical queries. The accelerator captures every domain event the platform emits (Entity, Workflow, Product, etc changes), enriches each event against the source APIs, and lands a structured, queryable data mirror in PostgreSQL. Financial institutions can then run their own reporting and analytics on top of a real-time, audited mirror without putting load on the platform's transactional path. Often using SnowFlake's PostgreSQL connector to import the data to in-house data lakes.

Both webhook (push) and polling (pull) ingestion are supported, so the system works against tenants that do not expose real-time event streams. Historical backfills and per-tenant reseeds are first-class operations rather than afterthoughts.

Architecture

Seven specialised .NET 8 microservices orchestrated with Aspire and deployed to Azure Container Apps. The pipeline is event-driven end to end: an Event Listener fronts the source platform's webhook and polling integrations, publishes onto RabbitMQ with dead-letter queues and idempotent retry handling, and an Event Handler consumes, enriches, and persists each message. Separate services own seeding & migration (Data Seeder API), workflow-driven ingestion (Flows API), native report syncing (Advanced Reporting Handler), and scheduled maintenance (Background Services).

The data store follows a Medallion architecture in Azure PostgreSQL Flexible Server. Raw event tables retain the original JSON payload and a complete audit trail, structured tables hold the relational projection, and materialised views provide query-optimised reporting surfaces that BI tools (Power BI, Tableau, custom SQL/API consumers) connect to directly. No proprietary semantic layer, no vendor lock-in on the read side.

Ingestion & enrichment

Webhook events are accepted, validated, and immediately published to queues so the source platform's delivery is acknowledged within milliseconds. The handler consumes asynchronously, enriches events through each data domain's Query API, and writes a normalised row alongside the original payload. Polling is symmetrical: the listener tracks watermarks per event type and replays gaps without producing duplicates downstream. The Flows API gives the source platform's workflow engine a direct ingress path at specific points in a given workflow, used when highly granular data is not a requirement.

Reporting layer

The Advanced Reporting Handler mirrors the source platform's native reporting engine: it syncs the catalogue of report queries, triggers scheduled executions, fetches outputs, and lands them in dedicated tables alongside the event-derived data. Background workers refresh materialised views on configurable cadences and manage retention. The result is a single PostgreSQL schema that combines event-sourced history, current-state projections, and pre-aggregated report outputs, all queryable with standard SQL.

Operations & deployment

Each tenant is a self-contained deployment. Clients fork the repository, populate tenant configuration, and run a fully automated GitHub Actions pipeline that provisions Azure infrastructure, builds and publishes container images, and brings the stack up in roughly fifteen minutes. Health endpoints, dependency probes, and Application Insights telemetry ship by default. The same artefacts run on-premises for clients that cannot deploy to Azure.