Structured market data
29 project records across 14 countries
The seed model covered solar, wind, storage, hydro, hydrogen and efficiency projects with country, stage, MW size, CAPEX, grid status and tags.
Live motion layer · Build pattern
Policy Copilot and Semantic Matching
Inputs, AI logic, review boundaries and public-safe outputs
Use case · Real build pattern, generalised
Policy copilot and semantic matching platform
Use approved source material, embeddings and weighted matching rules to answer market questions and connect opportunities to the right projects or partners.
Based on real implementation patterns from the ManchesterAI build archive, generalised for public explanation. This is not a named client claim, not a guaranteed metric and not regulated advice.
Clean-energy marketplace architecture proof29 project records · 15 investor mandates · 14 supplier profilesWeighted matching with explainable factors
Problem pattern
The work this kind of system removes
Teams working with dense policy, funding or market information often waste time reading long guidance documents, comparing eligibility rules and matching opportunities manually.
A good version does not replace judgement. It narrows the search space, shows why a match was suggested, and keeps approved source material visible.
Architecture proof
Not a chatbot demo: a marketplace-style decision system
The build pattern came from a clean-energy platform architecture where policy guidance, project records, investor mandates and supplier profiles sit in structured tables before AI is allowed to answer or match.
Matching candidates
15 investor mandates + 14 supplier profiles
Investor geography, technology focus and ticket-size fields were separated from project facts so matches can be explained rather than guessed.
Policy grounding
3 country policy briefs
The policy copilot uses approved brief fields such as incentives, permitting steps, lead times, agencies and source references before producing checklist-style guidance.
Vector-ready design
1536-dimensional embedding fields
The database includes pgvector-style embedding columns for projects, investors and suppliers, aligned to OpenAI text-embedding-3-small dimensions.
Public-safe proof: these are architecture and data-model facts from a real build pattern, not a claim that a named client achieved investor outcomes. The useful commercial promise is faster review, clearer shortlists and visible reasoning.
Transparent scoring
Matches should show why they were suggested
The matching layer combines semantic similarity with practical business filters, then returns the top matches with a factor-by-factor explanation and due-diligence disclaimer.
40% semantic fit
Embeddings compare the project story, investor thesis or supplier capabilities beyond exact keyword overlap.
25% geography fit
Country or regional coverage is scored separately so a strong text match does not hide a poor market fit.
25% technology fit
Solar, wind, storage, hydrogen and other technology focus areas stay visible in the score.
10% size/ticket fit
Project CAPEX or MW size is compared with investor ticket ranges so shortlist quality is not purely linguistic.
Model the records
Separate project, investor, supplier and policy fields before adding AI to the workflow.
Ground the answers
Keep approved sources, citations and disclaimers visible so the copilot supports judgement rather than replacing it.
Score the shortlist
Combine semantic fit with geography, technology and size/ticket rules, then show the reasoning.
Monitor usage
Log feature usage, response time, success/error state and review quality before expanding the tool.
Next step
Want to scope this as a controlled build?
Send one real workflow and the data it uses. We will map the safest first version and the review point before writing code.