Agora — feedback-analysis product

Agora takes a client's raw feedback — verbatim survey responses, reviews, open-text comments — through a preset analysis pipeline and into a live, login-protected dashboard: hierarchical themes, stance, and sentiment, with every AI-proposed coding decision visible and human-verifiable rather than taken on faith. It's built on Django with an Astro/Svelte frontend, and it's currently in productisation — the pipeline works, the product doesn't have paying customers yet.

1. Where it came from

Agora is a fork of the analysis-pipeline IP behind Stoa, the sovereign consultation-analysis platform I co-developed for a New Zealand qualitative-research firm (see the Stoa case study). Fridai retains that IP under our hybrid arrangement with the client, and Agora is the general-purpose product built on top of it — the same theme/stance/sentiment coding engine, aimed at feedback analysis rather than government consultation review.

2. Accuracy

Sentiment and polarity classification are benchmarked internally against SemEval-2014 gold data: 0.95 sentiment accuracyand 0.93 polarity accuracy across 800 gold-labeled sentences. That's the number I hold the pipeline to before trusting its output enough to put in front of a human reviewer.

3. Human-verifiable by design

The dashboard doesn't just display AI output — every theme, stance, and sentiment call a reviewer disagrees with can be corrected, and the hierarchy (categories down to leaf-level themes) means a reviewer can audit a high-level summary claim all the way down to the specific verbatims behind it. That review discipline is inherited directly from Stoa, where it was a hard requirement, not a nice-to-have.

4. Portfolio demo

Because Agora itself is a login-gated product with no paying customers to showcase yet, the public-facing demo is a separate, deliberately open artifact: a static case-study page replaying a real traced pipeline run over 1,643 public Emirates reviews pulled from Trustpilot and Skytrax, a captured walkthrough of the onboarding wizard, and a dashboard tour — plus a run-it-yourselfdocker compose up bundle that runs the real application (full dashboard, filtering, human-review flow) against that same pre-analysed Emirates dataset. The proprietary AI pipeline code itself is excluded from that bundle — a disclosed stub stands in for it — and the demo carries no login gate, unlike the production product.

I deployed the whole thing myself: a Hetzner VPS behind a Caddy edge with automatic TLS, per-engagement Docker Compose stacks, a hardened host (key-only SSH, ufw, fail2ban), uptime monitoring, and cost-capped LLM runs so a demo can't run up an open-ended bill.

5. Evidence

• Exhibit 11 — per-engagement VPS deploy (Caddy + Docker, multi-tenant isolation)
• Exhibit 12 — Docker image leak scanner (proves the pipeline stub, not the real code, ships in the public bundle)
• Capabilities matrix — every claim on this page traced to evidence