United Nations: Annual Report Automation & Knowledge Platform
Turn fragmented institutional knowledge into governed, reviewable reports and reusable AI access.
Context
United Nations country offices, where the evidence behind annual reporting is held across many separate systems and document repositories and had no reusable way to be queried or reused.
Constraint
Country-office knowledge and reporting inputs were fragmented, inconsistent, and often difficult to reuse. Annual reporting required substantial manual effort to find evidence, assemble a coherent narrative, and review the result across sources of uneven quality.
System
Built a Databricks-based ingestion pipeline that standardized heterogeneous source material, enriched it with metadata, and prepared it for grounded retrieval. The same knowledge layer was exposed through an MCP-compatible interface for reusable chatbot access. For annual reports, designed and built a LangGraph workflow with separate retrieval, drafting, LLM review, and targeted revision stages: sections that did not meet evidence, completeness, or structural checks were sent back with reviewer feedback rather than accepted unchanged.
Outcome
Automated annual-report production for United Nations country offices, removing a recurring manual reporting burden while retaining an evidence-grounded review loop. The underlying knowledge platform also created a single AI access layer for interactive questions and future knowledge-work automation.
Approach
- 01IngestionBuilt Databricks-based pipelines to bring inconsistent, heterogeneous source material into a common processing layer: extraction, normalization, metadata enrichment, chunking, and incremental updates.
- 02Grounded retrievalCombined semantic retrieval with document metadata and source context, then exposed the resulting RAG capability through an MCP-compatible interface so the client's chatbot could invoke it as a reusable tool rather than a one-off integration.
- 03Agentic report generationDesigned and built a LangGraph-based stateful workflow with distinct retrieval, drafting, evaluation, and revision stages to automate annual-report production for country offices.
- 04Automated evaluation & self-correctionScored generated sections against completeness, consistency with retrieved evidence, and structural requirements, routing failing sections back through a targeted regeneration loop using evaluator feedback.
- Agent API Integration - AI agent integration services for an existing agent and one authenticated business system. I implement the tool contract, permissions and failure handling, then test the connection against the actions your team has approved.
- Agentic Workflow Automation - Build a multi-step AI workflow around evidence, review and human decisions. I design and implement retrieval, drafting, evaluation and targeted revision for tasks such as report preparation, with explicit boundaries for any connected tools.
- AI Opportunity & Roadmap Assessment - Decide what to build, evaluate or connect before committing implementation budget. I review your use case, existing systems and data readiness, then set out the options, dependencies and a practical next scope.
- Chatbot Testing & Evaluation - Chatbot testing services for teams preparing a release or model change. I turn the questions and evidence your team approves into repeatable checks, a prioritized defect report and a regression runner you can keep using.
- Fractional Technical Leadership - Senior AI and data engineering guidance alongside your team. I help review architecture, resolve technical decisions and shape implementation priorities. We agree a regular review cadence and the decisions or implementation work I will own.
- Architecture & Implementation - Work directly with a senior engineer to deliver a defined AI or data project. I take responsibility for architecture, hands-on implementation and technical handover, using acceptance criteria agreed with your team.
- Private Knowledge Assistant Implementation - Turn an approved policy, SOP or document collection into a private knowledge assistant. My AI knowledge assistant development services cover ingestion, retrieval, source-linked answers and handover, with content and access rules approved by your team.
Discuss a similar initiative
30 minutes to map the decision, the data behind it, and the shortest credible path to a production outcome.