JN-66

JN-66 is an analysis droid from Star Wars: Attack of the Clones — a small, dome-shaped unit found in the Jedi Temple archives, built for quiet, methodical research. That felt right for a personal finance agent that lives in your terminal and does the number-crunching so you don't have to.

JN-66 is a self-hosted personal financial intelligence agent for households. Ask it natural-language questions about your spending, accounts, subscriptions, and transactions. It answers using a ReAct agent loop backed by any OpenAI API-compatible LLM — works with Ollama, OpenWebUI, or any hosted provider.

India-first: amounts in INR/paise, UPI/NACH/NEFT/IMPS payment modes, VPA-based counterparty identity.

Fully local. Zero data sharing. Designed to run entirely on your own hardware — your financial data stays in your local PostgreSQL instance and the LLM runs locally. Nothing leaves your network. Tested end-to-end on an RTX 3060 12 GB with multiple local LLMs via Ollama and OpenWebUI.

---

What it can do

• Spending breakdowns — "How much did I spend on food in April?"
• Transaction search — "Show me UPI payments above ₹2000 last month"
• Account summaries — savings, credit cards, wallets, loans — assets and liabilities
• Recurring payments — subscriptions, EMIs, UPI AutoPay, NACH mandates
• Label management — tag any transaction mid-conversation: "Label the Zomato one as food-delivery"
• Memory — tell it facts once, it remembers: "My Netflix ₹649 on HDFC CC every month is a subscription"
• Multi-user — knows who it's talking to; scopes data per user, supports household queries
• Fixed deposits — record FDs, track maturity dates, mark renewals (with per-term rate history) and closures
• Investments — Zerodha equity, SGB, and mutual fund holdings with P&L (requires Zerodha config)

See AGENT.md for the full usage guide and example questions.

---

Stack

Concern	Choice
Language	Go 1.26
Database	PostgreSQL 18 + pgvector
LLM	Any OpenAI API-compatible endpoint (Ollama, OpenWebUI, etc.)
SQL	sqlc — no ORM
Migrations	golang-migrate, embedded in binary
Config	koanf (YAML + env override)
CLI	chzyer/readline
HTTP	gorilla/mux
Money	BIGINT paise (INR × 100) — no floats

---

Quick start

# 1. Copy and edit config
cp config.yaml.example config.yaml
# edit config.yaml: set llm.base_url, users, api_keys

# 2. Start PostgreSQL 18 + pgvector
docker compose up -d

# 3. Run migrations
make migrate-up

# 4. Seed sample data (Alice + Bob, 3 accounts, ~40 transactions + memories with embeddings)
make seed

# 5. Build and run
make build
./bin/finagent --user alice

Ollama default LLM base URL is http://localhost:11434/v1. OpenWebUI or any hosted OpenAI-compatible provider works too.

---

Configuration

Copy config.yaml.example to config.yaml and fill in the values. All keys can also be set via environment variables using the FINAGENT_ prefix and __ as the level separator (e.g. FINAGENT_LLM__BASE_URL).

Reference

Key	Default	Description
database.url	—	PostgreSQL connection string
database.max_connections	10	pgxpool max connections
database.auto_migrate	true	Run pending migrations on startup
llm.base_url	—	OpenAI-compatible endpoint (e.g. http://localhost:11434/v1)
llm.api_key	""	API key (empty for local Ollama)
llm.routing.chat_model	—	Model for general conversation
llm.routing.analysis_model	—	Model for calculations and breakdowns
llm.routing.tagging_model	—	Model used by the enrichment pipeline
llm.routing.embed_model	—	Embedding model for semantic memory recall (e.g. nomic-embed-text)
llm.routing.summarize_model	—	Model for session title generation
agent.max_tool_rounds	20	Maximum ReAct loop iterations per message
agent.history_messages	20	Conversation turns kept in LLM context
channel.cli.default_user	—	Username used when --user flag is omitted
api.listen	:8082	HTTP server bind address
log.level	info	Log level: debug | info | warn | error
log.format	text	text (dev) or json (prod)
users[]	—	Users to create/update on startup (username, name, email, timezone)
api_keys.<username>	—	Bearer token for the HTTP API
zerodha.users[].username	—	Username to attach Zerodha credentials to
zerodha.users[].api_key	—	Kite Connect API key
zerodha.users[].api_secret	—	Kite Connect API secret

---

HTTP API

# Start in server mode
./bin/finagent --serve

All endpoints require Authorization: Bearer <api_key> (set per user in api_keys config).

Method	Path	Description
GET	/api/health	Liveness check — no auth required
POST	/api/chat	Send a message, get an agent response
POST	/api/accounts	Create an account
GET	/api/accounts	List accounts for the authenticated user
POST	/api/import	Import transactions from a bank statement file

Chat example:

curl -X POST http://localhost:8082/api/chat \
  -H 'Authorization: Bearer <api_key>' \
  -H 'Content-Type: application/json' \
  -d '{"text": "What accounts do I have?", "session_id": "optional-uuid"}'

---

Development

make generate   # regenerate sqlc types after schema/query changes
make fmt        # gofmt + goreturns
make build      # compile to bin/finagent
go test ./...   # unit tests (no database required)
make eval       # behavioural eval suite against real LLM + seeded DB

See CLAUDE.md for architecture details, conventions, and what's deferred to Phase 2.

---

Eval results

make eval runs two suites back-to-back against the real LLM and a seeded database. Useful flags:

Flag	Effect
--verbose	Print full LLM round traces for failed agent scenarios
--only-enrich	Run just the enrichment suite
--run <name>	Filter to scenarios whose name contains the substring
--compare "model_a,model_b"	Run both suites sequentially against two models and print a side-by-side comparison table

Model comparison results are logged in docs/evals.md.

Agent evals

Fixed natural-language prompts fired at the full ReAct agent. Assertions check which tools were called, in what order, and what the final response contains.

Scenario	What it checks
account_summary	Calls get_account_summary, output mentions account name
spending_breakdown	Calls get_spending_breakdown, output contains ₹ amount
investment_direct	Calls query_transactions for a specific month, finds SIP amount
transactions_list	Lists last N transactions, output contains correct counterparties
recurring_list	Calls list_recurring, output contains subscription name
remember_fact	Calls remember_fact to store a user-stated fact
recall_after_remember	Recalls a fact stored earlier in the same session
label_transaction	Lists transactions then calls manage_labels to tag one
fd_list	Calls list_fds, output contains rate and maturity year
fd_record	Calls manage_fd to create a new FD from natural language
fd_incomplete_prompts_for_details	Asks for missing FD details rather than assuming
max_rounds_respected	Handles an ambiguous query without exceeding the round limit
has_zerodha_account	Calls get_investment_summary, confirms Zerodha is mentioned
equity_summary	Calls get_investment_summary, output contains equity holdings
mf_summary	Calls get_mf_holdings, output contains MF details
portfolio_total	Calls get_investment_summary, output contains portfolio total

Latest: 16 / 16 passed

Enrichment evals

Raw transaction descriptions sent directly to the enrichment pipeline. Asserts the correct category_slug is returned. Covers the previously misclassified cases (credit card payments, bank charges, SIP) and golden-path categories.

Latest: 24 / 24 passed

model:    gemma4:12b-it-qat via Ollama
hardware: RTX 3060 12 GB VRAM
total:    ~6 min (40 cases, real LLM calls)

---

What's not here yet (Phase 2+)

• Physical assets (car, gold, property)
• Auto-fetch bank connectors (currently import is CSV/XLS upload only)
• Slack / Signal channels
• Tax assistance

---

Part of the R2-D2 household swarm

JN-66 is one agent in a larger multi-agent system built to manage a household end-to-end. The system is orchestrated by R2-D2 — a central manager agent that receives requests, decides which specialist to delegate to, and stitches results together into a coherent response. Each sub-agent owns a distinct domain; JN-66 owns personal finance.

Think of it as a household staff: R2-D2 is the chief of staff who fields every request and routes it to the right person — the finance analyst (JN-66), the calendar keeper, the grocery planner, and so on. No single agent needs to know everything; they each do one thing well, and the orchestrator holds the group together.

Code for R2-D2 and the other sub-agents will be open-sourced soon.