Client: Brighton Harwood
Brighton Harwood is a London-headquartered investment firm managing a diversified portfolio across commercial property, renewable energy infrastructure, healthcare technology, and growth-stage technology ventures. Over 15 years the firm had accumulated thousands of deal memos, board packs, due diligence reports, financial models, and post-investment performance data — spread across SharePoint libraries, email archives, PDF repositories, and legacy systems.
Their investment committee wanted to answer questions that no spreadsheet could: “Show me every renewable energy deal we evaluated between 2018 and 2023 where projected IRR exceeded 15% but we passed — and explain why we passed.” That kind of institutional memory was locked inside documents only a handful of senior partners could recall.
Veriland partnered with Brighton Harwood to build an AI-powered investment intelligence platform on Microsoft Azure — combining Azure AI Foundry orchestration, fine-tuned language models trained on the firm's own decision patterns, and a retrieval-augmented generation (RAG) architecture over their entire deal history. The platform is underpinned by Dynamics 365 Finance & Operations for fund accounting, investor reporting, and portfolio performance tracking.
Brighton Harwood's growth from a focused property investor into a multi-sector firm had created a rich but inaccessible body of institutional knowledge. The challenges were both technical and organisational.
Veriland delivered across three interconnected workstreams — each building on the last to create an AI platform that doesn't just search documents, but genuinely understands how Brighton Harwood thinks about investments.
The foundation was extracting structured data from 15 years of unstructured documents. Veriland deployed Azure AI Document Intelligence with custom-trained extraction models that understand the specific layouts and terminology of investment memos, term sheets, and due diligence reports. The pipeline processes PDFs, scanned documents, Word files, and Excel models — extracting deal metadata, financial figures, risk assessments, and decision rationale into a structured knowledge graph.
This wasn't simple OCR. Each document type required its own extraction template — a term sheet has a fundamentally different structure to an IC paper or a post-investment review. Veriland built and trained extraction models for each, with human-in-the-loop validation during the initial processing of the historical corpus.
Generic language models don't understand how an investment committee thinks. Veriland fine-tuned a model on Brighton Harwood's own corpus — thousands of IC papers, deal reviews, and partner commentary — teaching it the firm's specific vocabulary, risk frameworks, and decision patterns. When the model encounters “cap table complexity: high,” it understands this as a specific risk signal that historically correlates with longer due diligence timelines and higher deal mortality rates at Brighton Harwood.
The fine-tuning was performed on Azure OpenAI Service using a carefully curated training dataset. Veriland's data science team worked closely with Brighton Harwood's senior partners to validate that the model's interpretations matched the firm's actual decision logic — not just surface-level pattern matching, but genuine understanding of investment reasoning.
The intelligence platform is orchestrated through Azure AI Foundry, which manages the entire pipeline: document ingestion, chunking and embedding, vector storage in Azure AI Search, retrieval-augmented generation, and response synthesis. When an analyst asks “What renewable energy deals did we pass on in 2020–2023 and what were the common reasons?” the system retrieves relevant document chunks from the vector store, feeds them to the fine-tuned model with appropriate context, and generates a sourced, structured answer with citations back to the original documents.
Critically, the RAG architecture ensures the model's responses are grounded in Brighton Harwood's actual data — not hallucinated from general training knowledge. Every claim in a response links back to a specific document, page, and passage, giving the investment committee full traceability.
Beyond question-answering, Veriland built a deal similarity engine that automatically compares new investment opportunities against the entire historical portfolio. When a new deal enters the pipeline, the system generates a similarity score against past investments — flagging when an opportunity resembles previous winners (or previous mistakes). It surfaces patterns that even experienced partners might not consciously recall: a correlation between management team composition and post-investment EBITDA growth, or a sector-specific risk factor that historically predicted underperformance.
The AI platform is connected to Dynamics 365 Finance & Operations, which Veriland implemented as Brighton Harwood's fund accounting and portfolio management backbone. D365 FO handles multi-fund accounting, investor capital account tracking, fee calculations, distribution waterfalls, and regulatory reporting. The integration means the AI platform can correlate its deal analysis with actual financial performance — closing the loop between investment thesis and realised returns.
The platform transformed how Brighton Harwood sources, evaluates, and manages investments — turning 15 years of accumulated knowledge from a passive archive into an active competitive advantage.
“We've spent 15 years building institutional knowledge, but most of it was trapped in documents only three people could find. Veriland didn't just make it searchable — they built a system that actually understands how we think about deals. When a new opportunity lands, the platform tells us which past investments it resembles, what risks to watch for, and what questions we should be asking. It's like having a partner with perfect memory and zero ego.”
“The fine-tuning made all the difference. Off-the-shelf AI tools gave us generic answers. This platform understands our shorthand, our risk frameworks, our decision language. When I ask it about cap table complexity, it knows exactly what that means for our investment process — because it learned from our own IC papers.”
The platform's architecture reflects the complexity of turning unstructured investment data into reliable, auditable AI-powered insight. Every component was chosen for production resilience, regulatory compliance, and the specific demands of financial services workloads.
Ingestion begins with Azure AI Document Intelligence, configured with custom extraction models for each document type in Brighton Harwood's corpus. Investment memos, term sheets, board packs, and financial models each have distinct layouts and semantics — a single generic extractor would miss critical nuance. The pipeline normalises extracted data into a unified schema, enriches it with metadata (fund, sector, vintage year, deal stage), and writes structured records to Azure Cosmos DB while pushing document embeddings to the vector store.
The language model was fine-tuned on Azure OpenAI Service using a curated dataset of 3,200 investment committee papers, deal reviews, and partner commentary spanning Brighton Harwood's full operating history. The training data was carefully prepared to capture not just factual content but reasoning patterns — how partners weigh risk factors, how they articulate concerns, and how deal language evolves across the investment lifecycle. Veriland used a multi-stage validation process: automated evaluation against held-out test sets, followed by blind review sessions where senior partners assessed the model's outputs against their own judgments.
The retrieval-augmented generation layer uses Azure AI Search as the vector store, with hybrid search combining dense vector retrieval and keyword matching for optimal recall. Documents are chunked using a semantic-aware strategy that respects document structure — ensuring that a deal's risk assessment isn't split across chunks in a way that loses context. The orchestration layer in Azure AI Foundry manages query decomposition (breaking complex questions into sub-queries), retrieval, re-ranking, and response synthesis — with citation injection that traces every generated claim back to its source document and passage.
The platform runs within a private Azure virtual network with no public endpoints. All data is encrypted at rest (AES-256) and in transit (TLS 1.3). Access is controlled through Entra ID with role-based permissions — ensuring that fund-specific data is only accessible to authorised team members. Complete audit logs capture every query, retrieval, and generated response, providing the full traceability required for FCA compliance in AI-assisted investment analysis.
End-to-end orchestration of the AI pipeline — document ingestion, embedding generation, RAG retrieval, response synthesis, and model lifecycle management.
Fine-tuned language model trained on Brighton Harwood's investment corpus, deployed with enterprise security and compliance controls.
Custom-trained extraction models for investment memos, term sheets, board packs, and financial models — converting unstructured documents into structured knowledge.
Hybrid vector and keyword search over the entire deal corpus, powering the RAG architecture with semantic-aware retrieval and re-ranking.
Fund accounting, investor reporting, fee calculations, and multi-fund consolidation — integrated with the AI platform to correlate insights with financial outcomes.
Structured deal metadata store — fund, sector, vintage year, deal stage, and extracted financial metrics powering the deal similarity engine.
Role-based access control with fund-level data isolation, secrets management, and comprehensive audit logging for FCA compliance.
Managed integration layer connecting the AI platform, D365 FO, document repositories, and external data sources with reliable messaging and full observability.
