Surplus.ai — Codebase Visual Map

Built & working

Not yet built

Jonas would build

⚡

Tech Stack

SOLID

Next.js 16 React 19 TypeScript (strict) tRPC Tailwind Vercel AI SDK OpenRouter Zod

🔒 Strict TypeScript: no any, no as T, Zod validation at all boundaries. This is unusually disciplined for a prototype.

▼ UI Layer

🖥️

UI — Drawer-Based Interface

BUILT

📦 Products

👥 Segments

💰 Pricing Policy

📐 Price Models

⛓️ Value Chain

💎 Value Drivers

⚔️ Competitors

🔄 Alternatives

🧮 Costing

🤝 Deal Pricing

📡 Customer Signals

📂 Each section is a workspace tab. Users interact via drawers that trigger agent runs. Landing pages exist but the product is workspace-focused.

▼ tRPC (type-safe)

🧠

Agent System — 25 Agents

CORE

Agent Registry · lib/ai/

🔍 Product Discovery

Find and identify products

✨ Product Enrichment

Add features, specs, details

⚔️ Competitor Discovery

Find who competes

🕵️ Competitor Deep-Dive

Detailed intel on competitors

🔄 Alternative Discovery

Find substitute products

💰 Pricing Analysis

Analyze competitor pricing

👤 Buyer Import

Import buyer profiles

🏢 Company Profiling

Full company analysis

💎 Value Drivers

What drives customer value

✅ Feature Cross-Val

Validate features across sources

🎯 Lead Discovery

Find potential customers

📊 +14 more agents

Enrichment, analysis, signals

Agent Architecture

defineAgent({ name: "competitor-discovery", model: "haiku-4.5", // per-agent model selection tools: [webSearch, webFetch], // tools available systemPrompt: "...", outputSchema: z.object({...}), // Zod validated output }) // Registry: central Map<name, Agent> for all 25 agents // CLI runner: test any agent from terminal // JSONL logs: every run traced to .jsonl files // Streaming: NDJSON to UI for real-time progress

Models (via OpenRouter)

Claude Haiku 4.5 (default) Gemini Flash (some agents) Per-agent model selection

▼ Tools & External

🔧

Tools & Integrations

BUILT

🔍 Exa AI (web search)

🌐 Web Fetch

📂 Workspace Queries

💾 Result Caching

📝 JSONL Trace Writer

🖥️ CLI Test Runner

▼ Data Layer

💾

Data Storage

YAML FILES ONLY

⚠️ No database. All workspace data lives in YAML files. This works for a prototype but can't scale to multi-user.

▼ What's missing

🚧

Missing Infrastructure

NOT BUILT

🗄️ Database (no DB at all)

🔐 Auth / user system

👥 Multi-tenancy

💳 Payment / billing

▼ What Jonas would build

🔥

Jonas's Production Layer

TO BUILD

🗄️ DatabaseReplace YAML with real persistence

🔐 Auth SystemUser accounts, sessions, roles

👥 Multi-TenancyIsolated workspaces per org

💳 BillingStripe, usage metering, plans

⚡ Parallel AgentsSub-agent orchestration (Task 1)

📊 Ops & MonitoringOTel, dashboards, alerts

🎯 Task 1 for tomorrow: Parallel sub-agent orchestration. Run multiple agents concurrently using map-reduce + p-limit + Promise.allSettled. This is the highest-impact improvement to agent throughput.

🔀

Request Flow — End to End

User clicks "Enrich Product" in workspace │ ▼ UI (React 19) ──→ tRPC call ──→ Server │ ▼ Agent Registry.get("product-enrichment") │ ▼ ┌─ ToolLoopAgent ─────────────────────┐ │ 🤔 Think: "I need product specs" │ │ 🔧 Tool: exa.search("Product X") │ │ 👀 Observe: got 5 results │ │ 🤔 Think: "Need pricing too" │ │ 🔧 Tool: webFetch(competitor.com) │ │ 👀 Observe: got pricing page │ │ ✅ Done: complete enrichment │ └──────────────────────────────────────┘ │ ├──→ JSONL trace log (every step recorded) ├──→ NDJSON stream to UI (real-time updates) │ ▼ Results ──→ Save to YAML (future: database) │ ▼ UI updates workspace view ━━━ Tomorrow's Addition ━━━━━━━━━━━━━━━━━━ User clicks "Full Analysis" (runs ALL agents) │ ▼ Orchestrator splits into parallel tracks: ┌─ pLimit(5) ──────────────────────────┐ │ Promise.allSettled([ │ │ productAgent.run(), ← slot 1 │ │ competitorAgent.run(), ← slot 2 │ │ pricingAgent.run(), ← slot 3 │ │ buyerAgent.run(), ← slot 4 │ │ valueAgent.run(), ← slot 5 │ │ ...20 more waiting... ← queued │ │ ]) │ └───────────────────────────────────────┘ │ ▼ Reduce: combine all results into workspace