Nova AI Stack
The reusable technical foundation. Every product builds on this — nobody starts from zero.
§05 introduced Product Builder (orchestration). §06 defines what Product Builder orchestrates on top of — the actual stack layers, tools, and conventions that make Nova AI Ventures’ “next product starts at month two” promise (§02 Lesson 03) real.
1. What Nova AI Stack IS
🏗️ NOVA AI STACK · THE PLATFORM
Reusable technical foundation. Every product in the Nova AI Ventures portfolio builds on this foundation — nobody starts from zero.
Anchor: Google Cloud Ecosystem (primary) + Cloudflare (complementary)
- One primary vendor for integrated ML + enterprise platform (Google)
- Cloudflare = edge + networking + cost-effective storage (R2, zero egress fees)
- Gemini as co-pilot in the dev loop
- Enterprise-grade compliance out of the box
- Vertex AI RAG Engine native
- Generous free tiers for early-stage products
- Google Partner status — aspiration, not yet formalized, strong relationships in place; target Q3-Q4 2026 formalization
Stack ≠ Product Builder
- Stack = infrastructure (tools, services, runtime)
- Product Builder = orchestration (process, factory, workflow per §05)
- The Stack is general · Product Builder contextualizes it per-product
Nova AI Stack = Nova AI Ventures’ proprietary setup. We’ve invested substantial work into integrating these pieces (Layers 01-07 + MCP ecosystem + templates) into a coherent whole. Per §06 section 6, every product built on Nova AI Stack gets a one-off license on this proprietary configuration.
What Nova AI Stack is NOT
- ❌ Multi-cloud complexity (deliberately single-anchor focus early)
- ❌ Proprietary ML frameworks that don’t port
- ❌ Monolithic framework (everything composable via APIs/MCPs)
2. Two Levels of the Stack
🎺️ TWO LEVELS · L1 DEV · L2 PRODUCT
The Nova AI Stack serves two distinct audiences. Think of it as two slices through the same technical foundation. Level 1 = what Nova AI Ventures builds its own tools on (Product Builder itself, Management portal, internal agents). Level 2 = what every product built on Nova AI Stack gets (the standardized deployment profile that every new SPV inherits).
Level 1 · Nova AI Development Stack
What builds Product Builder and Nova AI Ventures’ internal tools.
| Layer | Components |
|---|---|
| Frontend | Next.js 14 · TypeScript · Tailwind CSS · shadcn/ui · React Query · Zustand |
| Backend | Express.js · TypeScript · REST APIs |
| Data | PostgreSQL · Prisma · Redis |
| Infra | Docker · Kubernetes · Cloudflare · Grafana / Prometheus |
| AI | Anthropic Claude (structured output + background agents) |
| Processes | Planning · execution · Claude-driven design · Stitch UX/UI · Product Builder workflow · wave-based delivery |
What Level 1 enables: the Product Builder itself (product-builder.nova-labs.ai) · the Management portal (management.nova-labs.ai) · all Nova AI Ventures-internal AI Management-Suite tools · the blueprint site you’re reading now.
Level 2 · Nova AI Product Stack
What every product built via Product Builder inherits by default.
Philosophy: “Every product should be accessible via chat.” Chat-first (Telegram / WhatsApp) with optional additional surfaces. AI native · RAG-ready · agent-ready · token-metered from day 1.
| Layer | Default | Alternatives |
|---|---|---|
| Chat Layer | Telegram + WhatsApp (defaults ON) | Slack · Discord (optional per product) |
| AI Layer | LiteLLM Proxy (model routing + caching) + MCP Server + Agent Orchestrator (auto-locked when any bot enabled) | per-provider fallbacks |
| RAG | Vertex AI RAG Engine | 6 alternatives (Pinecone · Weaviate · Qdrant · PGVector · Chroma · Milvus) |
| Backend | REST API + Swagger + Admin Panel | — |
| Frontend | shadcn/ui (Nova AI default) | 5 alternatives (MUI · Ant Design · Chakra · Mantine · Radix raw) |
| Data Analytics | Amplitude + Meta Pixel + GA4 (SnapSell pattern) | Mixpanel · PostHog |
| Processes | planning · execution · Claude design · Stitch UX/UI · Product Builder workflow | — |
What Level 2 enables: every SPV product — SnapSell · Lumen · Otherth · My Story Machine · Iris AI · any future build — inherits this full profile via Product Builder’s 7-stage pipeline (per §05 section 2).
Single source of truth: Nova AI Stack = Level 1 + Level 2 + all the proprietary integrations and templates that bind them together. The Platform License (§06 section 6 below) licenses this proprietary setup to Child SPVs.
3. The 7 Layers · bottom-up
🏗️ NOVA AI STACK · 7 LAYERS
Layer 01 — Platform & Hosting
foundationPrimary: Google Cloud Ecosystem
- Google Cloud Platform — compute, storage, networking, IAM
- Google AI Studio — model experimentation, prompt engineering, gen AI workbench
- Vertex AI — managed ML platform, model hosting, pipelines
- Firebase — auth, realtime DB, hosting for simple products
- Google Workspace — Docs, Sheets, Drive (integration surface for clients)
- Gemini — primary LLM + co-pilot in the dev loop
- Status: Google Partner aspiration → Q3-Q4 2026 formalization
Complementary: Cloudflare
- DNS + domain ecosystem —
*.nova-labs.ai+*.dev.nova-labs.ai - R2 object storage — S3-compatible, zero egress fees (product assets, user uploads, media)
- Workers (optional, per-product) — edge compute for latency-sensitive workloads
- CDN — global edge caching
Layer 02 — Backend Infrastructure + Observability
- Docker — containerization, reproducibility
- Kubernetes / Cloud Run — production orchestration
- Traefik — routing, domain management, TLS
- GCP Logging + Monitoring — infra observability
- Sentry — error tracking (live across mobile + API in SnapSell; standard across products)
- OpenTelemetry — observability standard, cross-service
Layer 03 — LLM Gateway
Central role in the entire stack. Every product has a dedicated LiteLLM Proxy instance (not shared).
| Component | Function |
|---|---|
| LiteLLM Proxy | Single entry point for all AI calls + all API calls (not only LLM) |
| Pass-through cost | Nova AI Ventures sees the real cost from the provider · applies margin (30-50%) · client pays cost + margin |
| Per-product instance | Every product has its own LiteLLM deployment (isolation · scaling · custom routing) |
| Multi-model routing | Primary model → fallback → local fallback (Llama) |
| Caching layer | Cost optimization for repeated calls |
| Guardrails | Rate limits · PII scrubbing · content policies per product |
What it meters: input tokens · output tokens · model used · latency · cost (real + margin) · API call success/fail · per-product · per-SPV · per-user · timestamp.
Real-time dashboards in Layer 05 Admin UI.
Layer 04 — AI Orchestration
conductor layer- Nova AI (Layer 0 conductor) — cross-agent coordination, master orchestrator
- OpenClaw — agent session orchestration · watchdogs · cron · file edits · shell work
- PI (agentic execution) — bounded task execution engine
- Agent-to-agent (A2A) — direct agent communication protocol · multi-agent workflows
- LangGraph — stateful agent graphs · multi-step workflows · conditional routing
- Llama (+ local models) — on-premise fallback · privacy-sensitive workloads · cost optimization
- MCP Servers — a growing ecosystem (MCP Stitch, MCP Playwright, MCP GitHub, MCP Claude Code skills, etc.)
- Vertex AI RAG Engine — knowledge retrieval · long-term memory per product
Frameworks: pick per-product based on fit (LangChain, AutoGen, CrewAI — open choice)
Decision flow: task arrives → LangGraph decides flow → PI executes bounded tasks → OpenClaw orchestrates sessions → Nova AI L0 coordinates → MCP servers provide tools → RAG provides context
Layer 05 — Admin & Platform UI
Nova AI-level (cross-portfolio)
- Management portal (
management.nova-labs.ai) — token dashboards · per-product margin · deployment status · releases view · portfolio-wide view - Product Builder UI (
product-builder.nova-labs.ai) — product creation orchestration (7-stage pipeline)
Per-product
- Product Admin Panel — dedicated admin interface for the application owner (operator, SPV owner). Shows: users · usage · revenue · token consumption · deployment status · backlog · customer support
- Pre-setup wizards — Product Builder scaffolds an appropriate admin based on product type:
- B2B with workspaces — multi-tenant · workspace management · team roles · billing per workspace
- B2C — single-user · individual billing · simpler onboarding
Layer 06 — Product Surfaces
user-facingUI framework standard: shadcn/ui + Tailwind CSS + Next.js — baseline across all web products (Lumen tech stack pattern).
Surfaces (complete list)
| Surface | Tech | Notes |
|---|---|---|
| Web | Next.js + shadcn/ui + Tailwind | Default for B2B/B2C web apps |
| Conversational | Telegram bots · WhatsApp · Slack | Telegram is the primary runtime target — many products can be primary-Telegram (with no web UI at all) |
| Voice | Agent-powered voice interfaces | — |
| Mobile | iOS · Android (native or React Native) | SnapSell lessons: IAP + Google Play Billing + Sentry |
| Transactional + marketing | — | |
| API | REST / GraphQL | Explicit surface · Line 04 AI Management Suite often API-primary |
Layer 07 — Data Analytics
top of stackPurpose: Measurement · attribution · product analytics · marketing measurement · user behavior tracking.
Components
- Product analytics — Mixpanel / Amplitude / PostHog (pick per product · default: Amplitude · consistent with SnapSell stack)
- Marketing analytics — Facebook SDK (Meta Pixel / Conversions API) · Google Ads SDK · TikTok Pixel (as applicable per product)
- Web analytics — GA4 · Cloudflare Analytics (default for
*.dev.nova-labs.aistaging sites) - Attribution — cross-platform attribution flows · UTM handling · referrer tracking
- Cohort analysis dashboards — exported to Management Nova AI portal (Layer 05)
- Data warehouse (later stage) — BigQuery (Google anchor) for portfolio-wide analytics
- Consent management — cookie banners · GDPR/CCPA compliance gates · tied to Layer 05 Admin UI
Why this is its own layer: Analytics is NOT product business logic (Layer 06 Surfaces) · NOT orchestration (Layer 04) · NOT backend (Layer 02). It’s a distinct measurement stratum that cuts across all product surfaces and feeds back into Layer 05 Admin UI dashboards. SnapSell’s analytics build (Mar 2026) validated this pattern — now standardized across all new products.
4. Token Economy
💸 TOKEN ECONOMY · PAY FOR VALUE
Resolves confusion once and for all. Token Economy is NOT:
- ❌ Crypto tokens
- ❌ Nova AI Units (NU) as currency
- ❌ Equity instruments
- ❌ Subscription tokens / feature-gate tokens
Token Economy IS: usage-based billing via LiteLLM Proxy metering.
Mechanism (repeated for clarity): LiteLLM Proxy meters every call → tracks input/output tokens + cost + timestamp + product/SPV/user → Nova AI Ventures applies margin (30-50% typical) → client pays monthly.
Renameable per product (UX layer)
| Product type | Token branding (example) |
|---|---|
| Cooking app | Cook Coins |
| E-commerce / SnapSell | Coins |
| Wealth management (Lumen) | Lumen Credits |
| AI Management Suite deploy | Management Credits |
| Generic fallback | Tokens or Credits |
Rule: Token naming = branding decision per product (Vincent / Iris decide in
product-branding/) · mechanism underneath = Nova AI Stack Layer 03 metering. The name is frontend, the mechanism is backend.
Applicable to: L02a/b Apps · L03 SMB · L04 AI Mgmt Suite
NOT applicable to: L01 Platform license (one-off) · L05 Venture R&D (project fees)
Token Economy · The LiteLLM Proxy Mechanism (how it actually works)
Everything routes through a dedicated LiteLLM Proxy instance per product:
- LLM calls — model inference (Claude Opus, GPT-5, Gemini, Llama, etc.)
- API calls — 3rd-party API calls (Vinted, OLX, Stripe, LinkedIn, etc.) routed through the proxy for centralized metering
- Internal tool calls — MCP server invocations counted for internal usage tracking
For every call the proxy records: timestamp · product · user · endpoint · input size · output size · latency · real provider cost · Nova AI Ventures margin · client-facing charge. All stored in the proxy’s metering DB · surfaced to the Management Nova AI portal (Layer 05) in real-time.
Pass-through cost mechanism
User Action
Triggers a call
"generate photo"
→
LiteLLM Proxy
Per-product instance
meter + route
→
Provider
Real cost
$0.05
→
Charge User
+ 40% margin
$0.07 in “Snap Coins”
- User action triggers a call (e.g., “generate listing photo for SnapSell”)
- Product app sends the call through its dedicated LiteLLM proxy instance
- Proxy routes to the appropriate provider (cost: e.g., $0.05 for image gen + GPT-4o call)
- Nova AI Ventures applies margin (e.g., 40% → $0.07 total)
- User’s token balance deducted ($0.07 in product-branded tokens · e.g., “Snap Coins”)
- Proxy records the event · updates margin dashboards · triggers any billing alerts
Optimization levers as the app grows
- Model routing — fallback from premium model (Claude Opus) to cheaper model (Haiku · Gemini Flash) based on request complexity · saves margin · maintained via LiteLLM routing rules
- Caching layer — identical requests in a short window served from cache · zero cost · 100% margin
- Batch processing — non-urgent calls queued and batched for cheaper tier pricing
- Per-product cost policies — limit expensive model usage per user tier (free users get Gemini Flash · paid users get Opus)
This is not theoretical. Every optimization already deployed on the LiteLLM proxy at
llm.snap-sell.app and llm.nova-labs.ai. Applied to every new product via Product Builder stage 3 (Infrastructure Setup per §05).
5. Technical Non-Negotiables · 9 Rules
🔒 NON-NEGOTIABLES · 9 RULES
Every product built on Nova AI Stack MUST follow:
| # | Rule | Why |
|---|---|---|
| 01 | API-first | Everything exposed through APIs · no monolithic frontends |
| 02 | MCP-first | Agent tool integrations via Model Context Protocol · reusable across products |
| 03 | Dockerized | Reproducible environments · no “works on my machine” |
| 04 | Observability from day 1 | Structured logging · metrics · tracing · OpenTelemetry standard |
| 05 | Security from day 1 | Auth · PII scrubbing · secrets management · HTTPS everywhere |
| 06 | Agentic-ready | Designed for AI agent interaction (tool descriptions, structured errors, idempotent ops) |
| 07 | Token-metered | ALL LLM calls via LiteLLM Proxy · no direct API keys in client code |
| 08 | Git-native | Backlog = GitHub issues · deployment tied to commits · audit trail |
| 09 | Google-anchored, no lock-in beyond | Use Google ecosystem · but avoid proprietary frameworks that don’t port |
6. Platform License Logic
📜 PLATFORM LICENSE · HOW IT FLOWS
Fundamental principle:
Nova AI Stack is our proprietary setup. We invest significant effort to make the stack work. Every application built on Nova AI Stack gets a one-off license for our proprietary configuration.
Nova AI Stack is our proprietary setup. We invest significant effort to make the stack work. Every application built on Nova AI Stack gets a one-off license for our proprietary configuration.
6.1 License properties
- One-off license per application — not recurring, not per-user, not per-seat
- Covers the entire proprietary configuration of Nova AI Ventures (Layers 01-07 + integrations + templates + MCP servers)
- Cut the umbilical cord — if someone buys out the application, Nova AI Ventures can “cut the umbilical cord” without any issue
- Buy-back option — Nova AI Ventures retains right of first refusal if acquirer wants out
- Improvements flow back — patterns, new MCP servers, new templates from the portfolio flow back to the platform (stack self-improves)
6.2 License states — depends on product profitability
| State | Application status | License treatment |
|---|---|---|
| Pre-profitable | Product in development / launch / pre-PMF | License = part of Nova AI Ventures’ investment in the startup. May be 0 PLN/year. Possible annual maintenance fee as a loan (deferred payment) |
| Profitable | Product generates sustainable revenue | License activates commercially — annual commercial license kicks in. Terms negotiated or pre-defined in SPV docs |
| Buyout (external acquirer) | External party acquires SPV | One-off license lump sum paid to Nova AI Ventures. Umbilical cord cut |
| Sell-back to Nova AI Ventures | Operator wants to exit, Nova AI Ventures buys back | Nova AI Ventures takes over · license reverts to internal |
6.3 Annual maintenance modes (open menu · pick per deal)
| Mode | Description | Typical fit |
|---|---|---|
| Mode 1 | 0 PLN/year (pure Nova AI Ventures investment) | Early-stage · pre-PMF |
| Mode 2 (default) | Deferred loan (accumulates, paid at profitability or buyout) | Nova AI Ventures investment stance |
| Mode 3 | Fixed small retainer (~1,000 PLN/year) | Symbolic, formal-SPV status |
| Mode 4 | Commercial rate (negotiated) | Post-profitability transition |
Per-product decision: Brian (Finance) + Dora (Legal) set in SPV docs at stage G6. Default = Mode 2 (deferred loan) until profitability trigger.
6.4 Profitability trigger (contract-customizable default)
- 50,000 PLN MRR (recurring monthly revenue)
- + 3 consecutive months of positive contribution margin
When both criteria are met → license state transitions from Mode 2 deferred loan to Mode 4 commercial terms.
7. Stack Ownership by AI Management-Suite
🤖 STACK OWNERSHIP · WHO OWNS WHAT
Each of the technical agents in the AI Management-Suite owns specific parts of the Nova AI Stack. Their tooling, briefing packages, and responsibility domains map to layer names (not layer numbers, which are internal labels only). Non-technical agents (executive operations like Klara and Nina, and growth/BD like Aurelia) are excluded here — they interact with external systems (Gmail, Calendar, LinkedIn, CRM) rather than the Stack itself. For their roles see §07 AI Management-Suite.
| Agent | Role | Primary stack layers they own | Primary tool set |
|---|---|---|---|
| Ada | Head of Architecture | All layers (architectural authority) | Codebase analysis · Git-aware reasoning · architecture decision records |
| Dexter | Head of Software Engineering | Backend Infrastructure · AI Orchestration · Product Surfaces | PI (bounded execution) · OpenClaw (orchestration) · GitHub · Docker · Traefik |
| Vera | Head of QA | Product Surfaces (browser-facing) · Backend (runtime probes) | Playwright MCP (full click-through testing · regression suites · smoke tests) |
| Mira | Head of Product | Admin & Platform UI · Data Analytics (dashboards) | Product analytics tools · roadmap management · feature prioritization |
| Iris | Head of Marketing | Admin & Platform UI · Product Surfaces (email, social) · Data Analytics (attribution, cohorts) | Brand design (MCP Stitch, Claude Code) · content drafting · ad creative · campaign tools |
| Brian | Head of Finance | Admin & Platform UI · LLM Gateway (cost metering) · Platform & Hosting (GCP billing) | Financial modeling · token economy dashboards · margin analysis |
| Dora | Head of Legal | LLM Gateway (license mechanics) · Admin & Platform UI (compliance) | Legal template library · contract review · per-product compliance tracking |
Explicit ownership flags:
- Dora — Platform License legal structure · SPV agreements · IP assignments · per-product compliance
- Brian — Token economy pricing · margin per product · cost metering · SPV financial modeling · buyout calculations
Full AI Management-Suite roster (including executive assistants Klara · Nina · Aurelia, and Nova AI L0 conductor, and human engineer Petro Polishchuk) is in §07 AI Management-Suite.