Alpha Grove Strategies | Elliot Gluck

Alpha Grove Strategies — Qualitative social and political research powered by conversational AI.

Overview

Alpha Grove explores how AI can be used to collect, analyze, and model social behavior — with an initial focus on political trends and public opinion.

Opinion data is gathered via primary-source research at scale using our home-grown conversational AI platform, AGS-1 .

Landing site + app integration

The marketing site served as the first impression and primary entry point into the AGS-1 demo. A key requirement was making the landing site (built in Webflow) and the application (built in Next.js) feel like one cohesive product, especially across domain boundaries.

Managing session persistence (Webflow → Next.js)

Because users could enter the AGS-1 demo from nearly any page, I preserved “where they came from” so the app could render accurate back links and avoid dropping users onto a generic home page.

Captured the current marketing URL in Webflow
Base64-encoded it (small payload, safe in a query string)
Appended it as a session param on links into the app

// Webflow custom code (conceptual)
const currentUrl = window.location.href;
const encodedUrl = btoa(currentUrl);

const url = new URL(button.dataset.href);
url.searchParams.set("session", encodedUrl);

button.href = url.toString();

I chose a URL param over session storage because the data volume was tiny and cross-subdomain/session behavior can be inconsistent across browsers.

On the Next.js side, the app decoded the parameter (when present) and hydrated navigation state accordingly:

// AGS-1 (Next.js)
import { useSearchParams } from "next/navigation";

const searchParams = useSearchParams();
const sessionParam = searchParams.get("session");

const defaultBackUrl = "https://www.alpha-grove.com";

let backUrl = defaultBackUrl;
if (sessionParam) {
	try {
		backUrl = atob(sessionParam);
	} catch {
		backUrl = defaultBackUrl;
	}
}

Post-demo personalization (legacy v1 → v2)

The v1 demo began with voice-only onboarding and captured a user’s name to personalize the conversation context. While v2 moved to a more traditional chat UX, a remnant of the personalization flow remained on the Webflow side: a post-demo landing page that could hydrate “thank you” UI based on a base64-encoded demo parameter.

Encoded a single user-provided string ( demo ) as base64
Decoded and normalized capitalization before hydration
Replaced any element marked with demo="true"

// Webflow custom code (conceptual)
function toProperCase(str) {
	return str.replace(/\w\S*/g, (txt) =>
		txt.charAt(0).toUpperCase() + txt.substr(1).toLowerCase(),
	);
}

const demoParam = new URLSearchParams(window.location.search).get("demo");
if (demoParam) {
	try {
		const decoded = atob(demoParam);
		const formatted = toProperCase(decoded);

		document.querySelectorAll('[demo="true"]').forEach((el) => {
			el.innerHTML = `${formatted} —<br><span class="font-opacity-low">thank you for trying AGS-1.</span>`;
		});
	} catch (e) {
		console.error("Error decoding demo param", e);
	}
}

AGS-1 application (real-time conversational AI)

AGS-1 is a real-time conversational AI platform built for low latency and global availability.

Architecture

Frontend : Next.js application responsible for UI, user/session state, and routing across chat/voice experiences.
Backend : Express.js service that manages orchestration, tool routing, and conversation memory.
Transport : WebSocket-based realtime link between client and backend for streaming messages and low-latency interaction.

Edge deployment + stateful sessions

To get edge latency without giving up state, the backend ran on Cloudflare Workers with Durable Objects :

Edge-first request handling close to the user
Per-conversation state and memory stored in Durable Objects
Horizontally scalable concurrency with predictable session affinity This made it possible to handle extremely high concurrent conversation counts (chat or voice) with negligible latency, while keeping conversation memory consistent.

Multi-model orchestration (demo)

The AGS-1 demo used multiple model providers, selected by domain fit:

xAI Grok for real-time retrieval
Anthropic Claude for language understanding and response quality
Google Gemini for conversation analysis + embedding generation
ElevenLabs for voice

Conversation analysis (current focus)

A major ongoing area of development was conversation analysis : extracting structured beliefs/opinions from natural language and producing semantic knowledge graphs.

Generate nodes/edges representing stated beliefs and causal claims
Attach embeddings to graph entities to enable similarity search and clustering
Query across users to find uncommon correlations and latent drivers behind opinion patterns