Skip to main content

​
The Thesis

Commerce is shifting from search-and-browse to agent-mediated. A user tells their agent what they want. The agent discovers options, evaluates them against the user’s declared preferences, and executes a purchase — all with the user’s trust. That trust requires three things traditional commerce infrastructure doesn’t provide:
  1. Transparent product intelligence — Agents need structured, scored product attributes with explicit provenance. They need to know why a product is recommended, not just that it’s recommended.
  2. User-owned intent profiles — Preferences, constraints, and avoidances that the user controls and the agent consults — never sold to advertisers.
  3. Machine-native payments — Agents need to pay for things programmatically, without card forms or browser sessions. USDC over HTTP 402 gives agents a payment rail that works like an API call.
Podium provides all three.

​
Architecture

​
The Two-Lane Provenance Model

Every signal in Podium carries an explicit SignalSource:
SourceOriginWeight in Feed
USER_DECLAREDCampaign votes, quiz answers, explicit preferences70%
MARKET_DERIVEDEnrichment pipeline — product reviews, structured APIs, community data30%
BASELINE_INFERREDComputed aggregates from enrichment signalsUsed in attribute baselines
This is a deliberate architectural choice. Enrichment data is not synthetic intent — it’s structured market intelligence that makes the platform useful during cold-start while first-party user data grows organically. As users interact (vote in campaigns, rank products, purchase), USER_DECLARED signals naturally dominate the intent score. Market-derived data gracefully steps back to a supplementary validation role. Agents consuming the product feed receive provenance metadata on every item: 
{
  "intentScore": 0.42,
  "intentProvenance": {
    "USER_DECLARED": { "voteCount": 12, "weight": 0.7 },
    "MARKET_DERIVED": { "signalCount": 87, "weight": 0.3 }
  }
}
The agent — and the user — always know what’s driving a recommendation.

​
Key Components

Enrichment Pipeline

Multi-source product intelligence: ingest, extract, resolve, normalize, baseline. The data engine behind informed agent decisions.

Agentic Product Feed

The endpoint agents call to discover products — scored, provenanced, and filterable. No auth required.

x402 Payments

Machine-native USDC payments over HTTP 402. Agents pay for products (or API access) with a single fetch call.

Beauty Companion

A complete reference implementation: a personal shopping agent built on these primitives with Telegram, AI recommendations, and Privy wallets.

​
Companion API

The Companion API (/companion/*) provides generalized infrastructure for building personal agents:
EndpointPurpose
GET/POST/PATCH /companion/profile/{userId}Intent profile CRUD — preferences, constraints, avoidances
GET /companion/productsFilterable product catalog with category, brand, price range, and search
POST /companion/interactionsRecord typed interactions: RANK_UP, RANK_DOWN, SKIP, PURCHASED, PURCHASE_INTENT, NUDGE_OPENED
GET /companion/recommendations/{userId}AI-ranked product recommendations based on profile + interaction history
POST /companion/ordersCreate a concierge order (the agent handles fulfillment)
GET /companion/user/by-telegram/{telegramId}Link external identities to Podium users
The Companion API is vertical-agnostic. The Beauty Companion uses it for skincare; you could build the same pattern for fashion, food, supplements, or any product domain. The schema fields (skin type, concerns, etc.) are profile-specific — the infrastructure (interactions, recommendations, orders) is generic.

​
For Developers

If you’re building an agent on Podium, the typical integration path is:
1

Create user and intent profile

Use the Companion API to create a user and build their preference profile through conversational onboarding, quiz mechanics, or direct input.
2

Discover products

Query the agentic product feed for scored, provenanced product listings. Filter by category, price range, or attribute baselines.
3

Record interactions

As the user expresses preferences (likes, dislikes, skips), record interactions. These feed back into recommendation ranking and intent scoring.
4

Execute commerce

Create checkout sessions and pay via x402 (USDC) or Stripe. The concierge order model handles fulfillment on the user’s behalf.
5

Earn trust over time

The more the user interacts, the better the recommendations. First-party signals naturally overtake market-derived data in the intent score.