Wally - Your Wealth Ally

An AI wealth agent that filters market noise and tells you what actually matters to your portfolio.

Comment

Wally — Your Wealth Ally

Inspiration

Most finance apps are built to show you more.

More prices. More charts. More alerts. More headlines.

But when something actually happens in the market, the harder question is:

Does this matter to me?

That is the question behind Wally.

Instead of building another portfolio chatbot or trading assistant, we wanted to build an agentic financial attention system — one that understands a user’s portfolio, watches what is happening through the financial day, and decides what actually deserves attention.

What it does

Wally combines:

  • portfolio context
  • market movements
  • current news
  • sector exposure
  • financial-day events
  • user preferences

and runs relevant signals through an attention pipeline:

Observe → Filter → Match → Exposure → Materiality → Score → Surface

A headline does not automatically become an alert.

Wally first asks:

  • Does the user actually own the company?
  • How large is their exposure?
  • Is there additional sector exposure?
  • Is the movement unusual compared with the market or sector?
  • Is there enough evidence to investigate?
  • Should this be ignored, monitored, investigated, or surfaced?

A proactive example

In our demo scenario, HDFC Bank moves sharply.

Wally detects the event without the user asking, compares the stock move with the financial sector, calculates the user’s portfolio exposure, determines that the event has high portfolio-specific relevance, and opens an active financial case.

The user can then:

  • understand why the alert matters
  • research the event deeper
  • continue the same context in Copilot
  • call their wealth agent through voice
  • send the case for advisor review
  • carry the case into the evening financial wrap

The same context follows the user throughout the financial day.

Built around Google ADK

At the centre of Wally is a Google ADK TaskMaster agent.

TaskMaster acts as the root supervisor and coordinates specialist capabilities including:

  • Portfolio Agent — portfolio holdings, allocation and exposure
  • Market Intelligence Agent — portfolio-relevant market and news context
  • Research Agent — deeper source-backed investigation
  • Media Agent — financial briefs, voice and recap experiences

We use Gemini through the Google GenAI SDK / Vertex AI for agent reasoning, explanations and conversations.

Google Search grounding gives our market intelligence and research workflows access to current, source-aware information.

Firestore persists financial cases, conversations, financial-day state, user preferences and onboarding context.

For real-time voice, LiveKit connects users to the same TaskMaster intelligence used by text Copilot — voice is not a separate chatbot.

More than a chatbot

Wally is designed around a financial day rather than a single chat session.

Its workflow can include:

Morning Pulse → Portfolio Health → Market Watch → Event Investigation → Market Close → Evening Wrap → Tomorrow Preparation

Scheduled checkpoints and unexpected market events both feed into the same financial-day state.

That means the agent can remember:

  • what it noticed earlier
  • which cases are still open
  • what the user already asked
  • what needs follow-up later
  • what should carry into tomorrow

Personalized from onboarding

Wally also includes context-aware onboarding.

A user can provide a small amount of information such as:

  • age range
  • income range
  • investing experience
  • financial goals
  • preferred level of detail
  • alert sensitivity

Wally then suggests defaults for:

  • financial profile
  • goals
  • risk profile
  • alert preferences
  • checkpoint schedule
  • voice style

All suggestions remain fully editable.

The final user-selected profile changes how the agent behaves throughout the day.

Thinking about what might matter next

Wally does not stop at explaining what already happened.

For relevant but uncertain developments, it can create scenarios to monitor.

These are explicitly not predictions.

Instead, the system can reason about plausible positive, neutral and risk scenarios, identify what evidence would make each more relevant, and create future watch events.

The goal is to move from:

“What happened?”

to:

“What should I keep an eye on next?”

Voice that already knows the context

Users can also call their wealth agent.

Through LiveKit, the voice experience receives the same:

  • portfolio
  • active cases
  • relevant stories
  • timeline
  • conversation context

as the text Copilot.

So a user can naturally ask:

“What deserves my attention right now?”

“Why does the HDFC alert matter?”

“What changed since this morning?”

without starting from zero every time.

Safety by design

One of the most important decisions we made was separating deterministic financial-materiality logic from generative AI.

Rules score materiality. AI explains the context. The user decides.

Wally does not:

  • place trades
  • issue buy/sell/hold commands
  • guarantee financial outcomes

It provides information prioritization, portfolio context and research support.

Challenges we ran into

The hardest part was not generating answers.

It was building continuity.

We had to make proactive events, scheduled checkpoints, portfolio calculations, research, chat and voice all work from the same source of truth.

We also had to make sure that:

  • portfolio exposure calculations stayed deterministic
  • voice did not become a separate chatbot
  • agent context persisted across the financial day
  • uncertain news stayed clearly uncertain
  • the system did not cross into investment recommendations

What we learned

Building Wally changed how we think about financial AI.

The interesting opportunity may not be another assistant that answers more questions.

It may be an agent that knows when a question is worth asking in the first place.

Wally is our attempt at that:

less financial noise, more portfolio-specific attention.

Built With

Share this project:

Updates

Submission history