Turn account evidence into qualified opportunities, reviewed outreach, and owned next actions.
中文项目说明 · Demo script · Architecture · Technical reference
One operating path from account evidence to the next commercial decision.
Navo is an account-intelligence and outbound-orchestration workspace for industrial B2B teams.
It starts with a target account, gathers source-backed evidence, separates facts from inference, qualifies the opportunity, prepares a message for human review, and turns the reply into an owned next action. The workspace keeps that operating loop visible from the first account review to the next owned action.
Accounts → Signals / Evidence → Research → Qualification → Contacts
→ Mission / Play → Message / Sequence → Replies → Next Best Action
Model output is structured, evidence is traceable, workflow state is durable, and external action stops at a human checkpoint.
The product is organized around one operator journey, with each layer assigned a clear responsibility.
- Product scope — centers the primary experience on one operator loop with a clear beginning, decision points, and outcome.
- Workflow architecture — defined which steps belong to deterministic code, which decisions can use a model, and where execution must stop for review.
- Evidence model — separated source facts, external signals, qualification, recommendations, and human decisions so each claim has a visible origin.
- Execution controls — enforces iteration, account, continuation, idempotency, and outbound-action limits.
- Validation coverage — combines deterministic tests, browser coverage, a real-provider smoke path, and regression cases for entity consistency and evidence grounding.
- Interface system — keeps account, mission, approval, inbox, and analytics work consistent across one dense workspace.
I led the product definition, system planning, task decomposition, acceptance criteria, and trade-off decisions across product, workflow, data, model integration, and validation.
| Account intelligence | Mission workbench |
|---|---|
|
 |
| Source evidence and qualification stay beside the account. | A natural-language objective becomes a bounded, inspectable plan. |
| Human checkpoint | Reply loop |
|---|---|
 |
 |
| Proposed outreach is editable before approval and read-only after the decision. | A reply becomes a classified conversation, one next action, and one owned task. |
| Current workspace analytics |
|---|
 |
| Current query-backed values show conversion, qualification mix, and source performance. |
flowchart LR
A[Account] --> B[Source evidence]
B --> C[Research and signals]
C --> D[Qualification]
D --> E[Mission plan]
E --> F[Draft message]
F --> G{Human review}
G -->|Approve| H[Controlled delivery record]
G -->|Revise| F
H --> I[Reply]
I --> J[Next best action]
J --> K[Owned task and account memory]
The model proposes bounded structured decisions. Code owns linear scheduling, state transitions, identity, persisted facts, limits, and postconditions. The reviewer owns the external-action boundary.
| Area | Current behavior |
|---|---|
| Account intelligence | Source evidence, signals, research, contacts, qualification, memory |
| Missions | Structured planning, bounded execution, continuation, pause/resume/cancel |
| Message review | Edit, request changes, approve, reject, recorded read-only state |
| Reply handling | Classification, summary, commitments, linked next action and task |
| Analytics | Current conversion, qualification mix, source performance |
| Reliability | Durable state, idempotency, immutable attempt history, deterministic fallback |
| Provider path | Deterministic mock mode plus workspace-scoped DeepSeek and OpenAI-compatible connections |
Requirements: Node.js, pnpm, Docker.
git clone https://github.com/shawliu998/Navo.git
cd Navo
pnpm run bootstrap
pnpm run dev:mockOpen http://localhost:3100, choose DACH industrial outreach, and follow the prepared demo
path. dev:mock uses deterministic local fixtures and does not require an API key.
To run a Mission with a real model:
- Start the normal runtime with
pnpm dev. - Open Settings → AI runtime.
- Choose DeepSeek or OpenAI-compatible, then enter the base URL, model, and API key.
- Save the configuration and run Test connection. The key is encrypted before storage and is never returned to the browser.
- Create a Mission. Web planning and worker execution now resolve the connected provider for the active workspace; Mock remains the fallback when no connection is active.
pnpm run bootstrap generates the local credential-encryption key and pnpm run doctor verifies
the install. pnpm smoke:ai runs one paid, end-to-end opportunity Mission with the tested workspace
connection and prints only non-secret acceptance evidence. For the full environment reference, see
technical reference.
| Check | Result |
|---|---|
| Lint | 7/7 workspace tasks passed |
| Typecheck | 7/7 workspace tasks passed |
| Tests | 192 passed; 13 configured integration tests skipped |
| Live provider smoke | DeepSeek discovery → outreach continuation completed with both plans in AI mode, no fallback, and one DRAFT |
| Product E2E | 15/15 Playwright scenarios passed |
| Focused UI checks | Inbox and Analytics passed with zero console errors or warnings |
| Core workspace QA | Account, Mission, Approval, Inbox, and Analytics flows verified |
The most useful review path is the 3–5 minute demo. Validation commands, runtime contracts, and environment details are in the technical reference.
apps/web Next.js product workspace
apps/worker Mission and reply execution
packages/agents Structured model adapter and agent contracts
packages/domain Business rules and workflow types
packages/db PostgreSQL schema and queries
packages/workflows Research, signals, qualification, and drafting
docs Demo, product story, and technical reference
artifacts Product screenshots and packaged release material