RReuters · 14 min ago
Fed signals two more cuts this year as inflation cools faster than expected
BullishMacroReliefNVDASPY
Selected work
ANS · Agentic News Signals
A news analyst that works all night, on a Mac Mini. A financial news intelligence platform for a professional trader: it reads the market around the clock, separates signal from noise and tests every idea against real prices. We designed and built it end to end, from the first screen to the models that run it.
ANS / Intelligence, on your terms Local inference
A private mind.
Your machine.
Custom software. Local AI.
One continuous flow from news to signal.
01 / Public inputs
01News & filings
02Market prices
03Social research
ANSLocal AI engine
Mac Mini / Ollama
02 / Checked output
Less noise.
More context.
Client
A professional trader (anonymous)
Role
Lobos Design: product design, UX/UI, full-stack development and AI engineering
- Timeline
- April 2026 → ongoing
- Deliverables
- Product design
- UX/UI
- Design system
- Web app
- Android PWA
- Telegram bot
- AI pipeline
- Paper-trading agent
- Platforms
- Desktop web app
- Android PWA
- Telegram bot
- Tools
- Python
- FastAPI
- SQLite
- React
- Local AI models (Ollama/Qwen)
- FinBERT
- Groq
- Grok
The challenge
A professional trader spent their days reading dozens of outlets, Telegram channels and social posts, and copying prices by hand into a spreadsheet. The problem was never a lack of news. It was the opposite: too much of it, all shouting at the same volume, with no way to tell what actually moves a stock from what only makes noise.
Their rule was simple: "if I buy ten newspapers and the same story is on all ten front pages, that's the theme." No tool on the market showed them that.
The solution
We built ANS around two questions instead of one score:
- Impact: how much does this story matter to the market?
- Retail: how loud will it be for ordinary people?
The gap between the two, the spread, is the product. A loud story that matters little is noise to step around; an important story almost nobody is talking about is an edge for whoever reads it first. The rest of the work was turning that into an interface a trader reads in seconds, on the desk and on the phone.
01 /Own the system
Your intelligence stays close.
A custom AI platform, running on one Mac Mini in a private network.
Designed around one trader’s workflow, with local models, application logic and storage on the same machine. Privacy is an architectural decision: the core runs locally, while external integrations have an explicit role.
Private environment / one Mac Mini
Mac MiniApplication + models + database On device
02 / 04
Local inference
Ollama · Qwen · FinBERTClassification, extraction and scoring run on models hosted on the Mac Mini. The core intelligence runs next to the application and its database.
Almost all model calls are local. Selected cloud integrations remain available for specific tasks.
External connections, made explicit. News and prices arrive from outside. Optional Groq / Grok integrations and Telegram connect to external services. Core inference and storage run locally; social research has a $10/month cloud cap.
Local first02 /The product idea
Loud is not the same as important.
Two questions. Two scores. The difference is the signal.
How much does a story matter to the market? How much attention will it attract? ANS keeps these questions separate. Explore how the spread changes when importance and attention move apart.
Try the scoring logic / illustrative valuesMove the noise.
Move the noise.
Find the signal.
Spread / retail − impact+14
Noise outweighs impact.
A loud story is not automatically an important one.
03 /Under the hood
A headline becomes a decision.
Six connected stages. A clear responsibility at every step.
From collecting public sources to delivering a checked briefing, each stage does one job. Follow the story through the system to see where models contribute, where code takes over and where the trader stays in control.
01 / Public inputs
Many sources. One intake.
News wires, filings, market prices, X and Telegram enter a shared pipeline. The original source remains attached to each story.
01News wires02Filings03Prices04X / Telegram
Incoming articles + source metadata
04 /Explore the interface
Built to be read in seconds.
One workspace. A complete reading of the market.
Explore the actual product on desktop and phone. Attention puts the day in context, Track replaces the manual spreadsheet, and Signal Lab tests ideas against what prices did next. The trading agent uses simulated money; account figures are redacted.
01 / Product walkthrough
The market, in focus.
Stories grouped by event. Sources, coverage and social context in the same view.
The question it answersWhat is everyone talking about?
05 /From the design file
Designed down to the detail.
A visual language for decisions, not just dashboards.
The Pencil design explores how to make dense information readable: three scores on a news card, price and news on the same timeline, a table for individual sessions and scoring rules that stay visible. Explore these original UI studies below.
NewsCard / ScoreBars / TooltipInteractive design specimen
Sample headline and values from the design file. Select a score to read its meaning.
Retail reach · 0–100
How loud will this story be?
How loud this story will be for ordinary people. Above 50 the crowd is talking; spread = retail − impact tells you whether that noise matters.
Control Panel ScoringTrack / Control Panel
The price, with its context.
Price action, volume spikes and news days share one timeline. Toggle the chart layers to see how each carries a different kind of information.
NVDANVIDIA Corp.
Design sample / 19 Aug 2026Last close187,44+3,89 · +2,12%
Volume vs median1,9×48,2M
3-month range96%172,40 → 188,10
EarningsAug 27after close
6M / 19 Feb → 19 Aug
Swipe sideways to inspect the full chart.
From ANS UI v2.pen · sample design data, not live market dataToggle a layer to explore the chart
A common component language connects the product: rounded widgets, segmented controls, compact score bars and a consistent colour for each kind of information. The original design frames remain available beside the interactive versions.
06 /Engineering trust
The model proposes. Code checks.
Every generated text passes through validation before it reaches the trader.
An audit found moved numbers, invented companies and translation errors. We built a common path for model-written prose: write, validate, retry once, then fall back to computed facts. For paper trades, code also decides position size, stops and orders within the trader’s limits.
Explore the guardrail / illustrative example
Trust is a process.
Change the model’s output. See what the code catches.
01 / Source facts
Company: Example Co.
Revenue: $12.4 million
02 / Model output
Example Co. reported revenue of $124 million.
03 / Code verdict
Blocked
$124m does not match the source value, $12.4m.
Rejected → one retry → computed facts if still invalid.07 /In operation
One machine. An ongoing advantage.
The trader retired the spreadsheet. The system keeps reading.
In production since April 2026. The figures below are the project’s 20 October 2026 snapshot: almost all model calls run locally, with a fixed cloud cap for social research. A private product that can keep evolving with the person who uses it.
Results
- Articles read and classified, from 29 sources
- 109,000+
- AI model calls, almost all on local models
- 511,000+
- AI calls a day on a single Mac Mini
- ~4,000
- Automated tests
- 2,400+
- Monthly cap on cloud spend for social research
- 10 USD
Product design → full-stack development → AI engineering
From the first screen
to the model behind it.
Want something like this?
Tell us what you have in mind. We reply within one to two working days.
Next project
Bella Ciao Pizzeria
