Perch: an AI Copilot for Animal Advocates (When Generic LLMs Aren’t Enough)

I used AI tools to help with phrasing/editing; all arguments and final content were reviewed by me.

As part of the Sentient Futures Project Incubator, I spent the last 8 weeks building Perch, an AI copilot for animal advocates. It's still in early development, but I’ve learned enough to share what we're building and ask for feedback.

The Problem: The "Movement Memory" Gap

Animal advocates increasingly rely on ChatGPT and similar tools for strategic guidance: "How do I design a corporate campaign?" "What's the evidence on plant-based defaults?" "How do I navigate this policy landscape?"

General-purpose LLMs certainly help, but can underperform where advocacy work is most fragile: factual grounding, source traceability, and domain-specific strategy. Critical movement knowledge is scattered across PDFs, private Slacks and forums, internal documents, and conference 1-1s, and advocates either spend too long searching for this information or are forced to go without it. 

For example, during the campaign for Denver’s 2024 citizen-led ballot measure to ban fur, activists reportedly faced significant backlash from parts of the fly-fishing community, which uses animal-derived materials for flies and lures (see article). If you ask ChatGPT for risks to prepare for in a local fur campaign, it may give plausible generic opposition, but it misses this concrete prior-campaign lesson.

Perch as a Specialized Copilot

Perch is our attempt to solve that gap. We built a retrieval-augmented generation (RAG) system that draws on a curated knowledge base of animal advocacy research, news, and reports. RAG allows us to ground LLM responses in specific sources rather than relying only on model pretraining, so claims are directly traceable to indexed sources and easily verifiable. 

With Perch, we aim to shift AI assistance for animal advocates from broad generalist reasoning toward tactical, movement-specific insight.

Current state:

• Initial corpus ingestion from core advocacy sources, including:
  • Faunalytics research
  • The Humane League reports
  • Academic journals on animal welfare, behavioral economics, food systems
  • Policy/legislative documents
  • Selected EA Forum content relevant to animal advocacy
• RAG pipeline with semantic chunking and custom retrieval/reranking in place. Tech stack: FastAPI, LangChain, Pinecone, gpt-5-mini, Next.js frontend
• System prompts optimized through iterative evaluation
• Current demo use case: college-campus vegan advocacy (strategy, messaging, implementation)

Known limitations:

• Coverage is still limited by corpus scope. So far, we’ve prioritized resources for our current demo use case and need broader coverage.
• We want to ingest knowledge from advocacy Slack channels (e.g. Hive, Sentient Futures), conference talks, and internal organizational documents, but we need clear consent and privacy protocols first.
• We’ve run automated evaluations of response quality, but haven’t yet measured whether this translates into faster or better real-world decision-making.

What We've Learned So Far

I ran a comparative benchmark against ChatGPT on campus vegan advocacy scenarios, scoring outputs for operational detail, evidence grounding, and accessibility. Early findings highlight where Perch performs well:

1. Actionability: Perch provides concrete next steps (timelines, KPIs, constraint analysis) that ChatGPT typically omits.
2. Source specificity: Perch retrieves and cites backing documents; ChatGPT often makes unsourced claims.

I also evaluated Perch in isolation using an LLM-as-judge rubric. In that Perch-only evaluation, advocacy context and actionability were consistent strengths. Through iterative system prompt engineering and expanded corpus coverage, scores for the Evidence Base dimension on the rubric improved from an average of 1.8 to 2.9 (out of 3) across our benchmark scenarios. 

This signals that Perch is already improving evidence-grounded actionability in animal advocacy contexts, at least in our benchmarks. 

How You Can Help

I’d love input from EA Forum readers:

1. Eval prompts: send us real questions you ask LLMs for animal advocacy work, and we'll benchmark Perch against them.
2. Problem feedback: Is this actually a problem advocates face? Are there better ways to solve it?
3. Corpus suggestions: Which datasets, reports, or resources should we prioritize?
4. Pilot testers: When we’re ready, we want to test with advocacy organizations and measure whether Perch improves decision quality and speed in real workflows.

Try it (rough early version): https://perch-ai.vercel.app/
This is at prototype quality; outputs can be wrong and should be independently verified.

For the technically curious, you can view the codebase here: https://github.com/Mycelium-tools/perch 

Looking forward to feedback.