The problem: watching a portfolio is a part-time job
Owning even ten stocks means a never-ending stream of headlines, press releases, SEC filings and price moves. To stay on top of it properly, someone has to check the sources all day, sort the few items that matter from the hundreds that don't, research the ones that do, weigh them against their own limits — and remember to follow up later. Most people either spend hours doing it, or don't do it at all.
MarketBell does that work for you, all day, and only asks when a decision genuinely needs you.
On its first afternoon live on Azure, MarketBell read 467 news items, set aside 425 with a reason for each, investigated 46 with Nemotron — and brought just 2 to me.
AI understands the market. Rules control the money.
Live demo: mktbell-staging-api…azurecontainerapps.io · Code: github.com/Anthony-hue25/Marketbell (MIT)
The work MarketBell does so you don't have to
For the demo portfolio (6 holdings, a watchlist and a discovery pool — 14 companies), here is MarketBell's daily workload:
| Every day, MarketBell… | So you don't have to… |
|---|---|
| Checks news for 14 companies every 5 minutes — about 4,000 company checks a day — plus press-release wires and SEC filings | keep refreshing feeds and filing sites |
| Reads and triages every new item, removing duplicates, syndicated copies, wrong-company matches and low-credibility commentary, and logs a plain-English reason for each one it sets aside | skim hundreds of headlines to find the few that matter |
| Investigates only what's material, with two independent Nemotron reviews that must quote their sources word for word | read the sources and argue both sides yourself |
| Checks every case against your limits — position size, sector exposure, cash, price freshness and evidence quality | do the maths before every decision |
| Reviews every holding and watchlist stock once a day after the close | run a daily portfolio check |
| Handles repeat decisions automatically under rules you've approved | answer the same question twice |
| Revisits every decision at 7, 30 and 90 days against later evidence | remember to follow up |
| Brings you only the few decisions that need your judgement | — |
In its first minutes live on Azure, MarketBell read 250 news items, set aside every routine one with a reason, and concluded that nothing needed me. Today's live tally: [READ] items read, [SET ASIDE] set aside, [INVESTIGATED] investigated, [NEEDS YOU] needed me.
What one investigation replaces
When something material happens, MarketBell does in about 10 seconds (for ≈ $0.002 of Nemotron calls) what would take a person a focused half-hour:
- Finds the story across sources and confirms it's the right company from a credible publisher
- Researches what it means, citing exact quotes from the evidence
- Gets a second, independent opinion that never sees the first — disagreement holds the case
- Checks your hard limits in code, every time
- Either handles it within your authority — preparing a verified research note — or brings it to you with a plain-English explanation
- Schedules the follow-ups and records every step in a tamper-evident history
Teach it once, it handles the next one
When MarketBell asks, you can Approve, Reject, or Correct & teach. A plain-language correction becomes a proposed, narrowly scoped, expiring rule that grants nothing until you save it. After that, every matching case is handled for you — one correction on Phillips 66, and MarketBell prepares the Valero note on its own. One decision removes a whole category of future questions. Rules can be paused, edited or revoked any time, and never reach back to earlier cases.
What still comes to you
Only what genuinely needs a human: a material risk to something you own, a new kind of opportunity no rule covers, evidence too thin or too old to act on automatically, or rules that conflict. Every case shows what was observed, what the AI interpreted, how fresh the evidence is, which rules applied, and whether anything was actually done — a proposal is never shown as done.
MarketBell never trades: no brokerage endpoint, no order route. Its only action is preparing a research note.
Order of authority: Hard policy > Your approved rules > Your context > Model recommendation
How I built it
- NVIDIA Nemotron 3 Super on Nebius Token Factory (
nvidia/nemotron-3-super-120b-a12b) in five roles: Research, blind Challenge, turning corrections into proposed rules, 7/30/90-day follow-ups, and "Ask MarketBell" answers drawn only from its records. Every answer must match a strict JSON schema with citations verified in code; anything else fails closed and is shown as Not verified. - Built to run all day cheaply: triage means the model is called only for material news; the second review is skipped (and recorded) when it can't change the outcome; daily AI budgets keep a separate allowance for visitors so monitoring is never starved.
- Python backend: deterministic policy engine, always-on worker with database leases, append-only evidence store, immutable decision snapshots, hash-chained audit history.
- Free data sources: Google News RSS, press-release wires, SEC EDGAR and Alpaca daily bars (read-only).
- Live on Azure Container Apps: API, background worker and nightly audit job; PostgreSQL with Entra-only sign-in; secrets in Key Vault via managed identity — no passwords anywhere; email alerts if monitoring stalls, Nemotron fails or spend climbs.
- Quality gates: 222 automated tests on SQLite and PostgreSQL, security and secret scans, a container vulnerability scan, and a production-style start-up check on every push.
Challenges I ran into
- Making autonomy safe. Governance tests surfaced two subtle cases — an edited rule reaching back to earlier cases, and old evidence triggering automatic action. Both are fixed and covered by tests.
- Making all-day monitoring affordable. Careful triage and early stopping keep Nemotron calls to the items that matter.
- Shipping on a free-trial cloud. I worked within regional and quota limits by building the image on GitHub and importing it into Azure.
Accomplishments that I'm proud of
- It genuinely works while you're away: monitoring, triage and Nemotron investigations run on Azure with no browser open.
- It gets less needy over time — safely. One approval can retire a whole class of questions, and it can never quietly gain authority.
- Honest by design: exact quotes, dated evidence, no invented confidence scores, and a finished workflow is never presented as proof of profit.
What I learned
The most valuable thing an agent can do is take real work off someone's plate — and decide well what not to bother them with. Pairing a capable open model like Nemotron with deterministic authority gives both: judgement where it helps, guarantees where they matter.
What's next for MarketBell
- Connecting real portfolios through read-only account links
- Reading full articles alongside headlines
- Mobile notifications for the rare "needs you" moment
- A running "time given back" measure for each investor
Built With
- ai-agent
- azure-container-apps
- fintech
- human-in-the-loop
- investing
- llm
- nebius
- nvidia-nemotron
- python
