Jarvis is a sophisticated AI assistant designed specifically for Independent Financial Advisors (IFAs). Built on a "Board of Specialists" architecture, Jarvis autonomously monitors client data, identifies risks, and drafts professional communications.
View Demo Videos and Slide Deck on Google Drive
| Component | URL |
|---|---|
| Frontend | Vercel Dashboard |
| Backend API | Railway API Docs |
- Jarvis Orchestrator: The central brain that manages delegation and identity synthesis.
- Atlas (RAG Specialist): Deep-dives into client files, emails, and transcripts.
- Emma (Paraplanner): Drafts suitability reports and advisor correspondence.
- Colin (Compliance): Ensures all outputs meet UK FCA regulations.
| Context | Model |
|---|---|
| Direct Chat (Jarvis) | openai:gpt-4.1 |
| Heartbeat & Cron Jobs | openai:gpt-5-nano |
Jarvis is built using LangChain Deep Agents, a framework for creating sophisticated AI agents with tool use, planning, and sub-agent delegation. This provides Jarvis with:
- Autonomous reasoning with multi-step planning
- Tool orchestration with built-in filesystem, web, and custom tools
- Sub-agent delegation for specialized tasks
Instead of a static prompt, Jarvis builds its personality and context dynamically at runtime by reading workspace files:
SOUL.md- Core personality and valuesIDENTITY.md- Professional identity as a financial advisor assistantUSER.md- Information about the advisor Jarvis servesHEARTBEAT.md- Instructions for autonomous background checks
This allows the advisor to customize Jarvis's behavior by simply editing markdown files.
Jarvis delegates specialized tasks to three expert sub-agents:
| Agent | Role | When Used |
|---|---|---|
| Atlas | RAG Specialist | Searches client documents, emails, and transcripts |
| Emma | Paraplanner | Drafts professional client communications |
| Colin | Compliance | Validates outputs against UK FCA regulations |
Each sub-agent has its own system prompt, tools, and communication style.
Skills are modular, file-based instruction sets that extend Jarvis's capabilities without modifying code. Located in ./skills/, each skill contains:
SKILL.md- Instructions and workflows- Supporting scripts and templates
Jarvis automatically loads and uses relevant skills based on the task context.
Jarvis can schedule its own future tasks using APScheduler. This enables:
- Periodic reviews - e.g., "Review client portfolios every Monday at 9am"
- Deadline reminders - e.g., "Remind me about Sarah's annual review in 3 days"
- Autonomous monitoring - Self-scheduled tasks that run without human intervention
The dashboard displays all scheduled tasks and allows manual cancellation.
Jarvis maintains persistent memory through workspace files:
memory/daily_log.md- Record of daily interactions and outcomesmemory/client_notes.md- Important client-specific observations
This allows Jarvis to recall past conversations and build context over time.
To save tokens and avoid unnecessary chatter, Jarvis uses special tokens:
NO_REPLY- When there's nothing meaningful to say (e.g., simple greetings)HEARTBEAT_OK- When a background check finds nothing urgent
The API filters these tokens so they never reach the user interface.
Client documents (emails, transcripts, policies) are embedded and stored in ChromaDB for semantic search. This enables:
- Context-aware retrieval - Find relevant documents based on meaning, not just keywords
- Cross-client analysis - Search across all clients for similar situations
- Efficient long-term memory - Store large document collections without token limits
Jarvis has access to both default and custom tools:
Default Tools (from Deep Agents):
- Filesystem operations (read, write, list, glob)
- Web browsing and search
- Planning and task management
Custom Tools:
| Tool | Purpose |
|---|---|
get_market_news |
Fetches live UK financial news via Tavily |
find_files_updated_after |
Detects recently modified workspace files |
add_cron_job |
Schedules future tasks |
remove_cron_job |
Cancels scheduled tasks |
list_cron_jobs |
Shows all scheduled tasks |
retrieve_context |
Vector search in ChromaDB (Atlas) |
search_uk_compliance |
FCA regulatory search (Colin) |
Test Jarvis's ability to scan his entire "book" for urgent matters.
- Launch the Dashboard: Run
cd frontend && npm run dev. - Open the Chat: Navigate to the "Chat" page.
- Ask the Query:
"Show me anything in the last 10 days that looks urgent across my book (emails and meeting notes)?"
- Behind the Scenes: Jarvis will scan
datasets/**for files modified between Jan 28 and Feb 08, 2026, and use Atlas to identify risks.
The sample langsmith trajectory: here
Test how Jarvis identifies new incoming data and alerts you autonomously.
- Prepare the Data: Locate the sample email in
sample/2026-02-01_ill_situation.txt. - Upload:
- Go to the "Clients" page in the dashboard.
- Select Gareth Cheeseman.
- Upload the sample file as an "Email Archive" document.
- The Result: Within 30 minutes (or on the next heartbeat), a notification will appear in the Dashboard:
"🚨 Jarvis Alert: Gareth Cheeseman has emailed regarding income protection policies due to illness..."
The sample langsmith trajectory: here
python scripts/ingest_documents.pyuv run uvicorn jarvis.api:app --reload --port 8000Note: The heartbeat scheduler now runs as a background thread within the API.
cd frontend && npm run devsrc/jarvis/: Core Agent and Sub-agent logic.workspace/: Operating environment, prompts, and client datasets.frontend/: Advisor dashboard (React + Vite).sample/: Standardized text data for system validation.skills/: Modular skill definitions.
Integrate Jarvis with WhatsApp, Telegram, and Slack gateways so advisors can interact from their preferred messaging platform - not just the web dashboard.
Connect Jarvis directly to Gmail, Outlook, and calendar APIs via MCP (Model Context Protocol) servers. This would enable:
- Automatic email ingestion without manual uploads
- Meeting scheduling and follow-up reminders
- Real-time client communication monitoring
Embed Jarvis in group chats between advisor and client families. The advisor could invoke Jarvis with @jarvis to:
- Answer client questions with compliance-checked responses
- Pull relevant policy details mid-conversation
- Draft follow-up actions in real-time
Evaluate Jarvis quality using:
- Human judges - Advisor ratings on response quality
- LLM-as-judge - Automated scoring for consistency and compliance
- Task completion metrics - Success rate on standard scenarios
Build a library of shareable skills:
- Annual review workflows
- Product recommendation templates
- Regulatory update handlers
- Client onboarding wizards
"I'm not just a chatbot; I'm a teammate who never sleeps." — Jarvis