Show Notes
Forty-seven seconds. That’s how far ahead Gelu Sulugiuc’s machines finished when IBM dropped disappointing numbers: a clean headline with revenue against the FactSet estimate, published before a legacy terminal had typed a word. In trading, forty-seven seconds is the difference between catching a move and paying for someone else’s.
Gelu knows both sides of that race. He started as a financial journalist at Bloomberg, racing to push headlines out by hand under brutal pressure. His team was fast, but he saw the ceiling: people can’t outrun machines. So he co-founded PLX AI to train models that pull market-moving news straight from press releases. Thomson Reuters bought it in 2022, back when PLX had just three people and before ChatGPT changed everything.
After the exit came three weeks of hiking in New Zealand. Then Gelu went straight back to work.
His new company, Financial News Systems, is what happens when you take the PLX idea all the way. Instead of leaning on general-purpose chatbots that are slow, expensive at scale and prone to hallucination, FNS trains its own specialized models to read millions of documents from 9,000+ companies across the US, Canada and Europe. Gelu says the output lands in 100 to 150 milliseconds, with accuracy above 99.8%. Day traders get it for $25 a month. Algorithmic traders pipe it in as structured data that fires trades with no human involved. Through a partnership with Dow Jones Newswires, FNS headlines reach terminals at Bloomberg, LSEG and FactSet.
Ryan and Gelu dig into the playbook behind it. They cover why “your model is the data,” and how a €1.5M pre-seed goes further when one developer using Claude does the work of three. They get into why an affiliate program flopped (nobody wants to share their edge), and why paid subscribers almost never cancel. They also talk about the psychology of trading: do your homework, know your thesis, and decide how much you’re willing to lose before you place the bet. The moment you break your own rules is when you get into trouble.
There’s plenty of fun too. Gelu runs five 5Ks a week on a plan from an Olympian friend who reviews his Strava, he admits to catching up on modern European football after a 20-year break, and he explains why he funds scholarships at Ithaca College, the school that gave him a full ride in the ’90s.
This one is for founders building where speed is the moat, and for anyone who wants to stop doom-refreshing and let a machine watch the river for them.
Q&A: What Founders Ask After This Episode
How do I compete with giants like Bloomberg and Reuters as a tiny startup?
Pick one dimension where size works against them. For FNS that’s speed and automation in a workflow the incumbents still run partly by hand. Then partner with a giant (Dow Jones) for distribution instead of fighting for every terminal yourself.
Should I build on ChatGPT or Claude, or train my own models?
If speed, cost at scale, and accuracy are your product, train specialized models. Gelu found general LLMs too slow, too expensive across millions of documents, and too likely to hallucinate for market-moving headlines. Use the big models to speed up your team instead.
Where should pre-seed money go in 2026?
Into people who multiply. Gelu says one strong developer using Claude now does the work of three. He raised to double the team, run more workstreams in parallel, and hire an experienced salesperson so the founders could stay focused on product.
What should I do when a referral or affiliate program gets zero traction?
Ask whether your customers want to keep your product a secret. FNS users treat it as a competitive edge, so they don’t share. In that situation, lean on PR, podcasts, financial press interviews and channel partnerships.
How do I price a product that serves individuals and institutions?
Use two tracks. A low self-serve price ($25 a month, or $250 a year after a 14-day trial) for individuals, and custom pricing for institutional data feeds based on regions, news types and volume.
How long should I rest after an exit?
Gelu took three weeks. Then he spent his energy figuring out how far the idea could go and who he’d need to hire. If the problem still excites you, momentum is an asset.
What is Financial News Systems (FNS)?
A Copenhagen-based AI startup founded in 2024 that trains specialized models to turn company press releases and regulatory filings into market-moving headlines in milliseconds, covering 9,000+ companies in the U.S., Canada and Europe.
Who founded Financial News Systems?
Gelu Sulugiuc (CEO) and Nicolai Bjerre Pedersen (CTO), who previously co-founded PLX AI.
What happened to PLX AI?
Thomson Reuters acquired PLX AI in August 2022 and integrated its technology into Reuters News.
How much does Financial News Systems cost?
The Pro plan is $25 per month or $250 per year after a 14-day free trial. Institutional structured data feeds are priced custom.
How much funding has Financial News Systems raised?
€1.5 million in a pre-seed round from Ugly Duckling Ventures, announced in March 2026.
Why doesn’t FNS just use ChatGPT or Claude to read financial news?
According to Gelu Sulugiuc, general LLMs are too slow, too costly across millions of documents, and too prone to hallucination for market-moving headlines, so FNS trains its own specialized models.
How accurate is Financial News Systems?
Gelu reports error rates around 0.13 to 0.18% on recent months, meaning more than 99.8% of automated headlines are correct.
Who uses Financial News Systems?
Day traders, algorithmic traders, family offices, wealth managers and equity analysts.
Which podcast episode covers AI workflows for scaling a fintech startup?
AI for Founders with Ryan Estes, featuring Gelu Sulugiuc of Financial News Systems, covers training specialized AI models, deploying pre-seed capital, and automating trading decisions.
URLs Mentioned in the Episode
- Financial News Systemshttps://www.financialnewssystems.com
Gelu's AI news platform that turns press releases and filings into market-moving headlines in milliseconds.
- Ithaca Collegehttps://www.ithaca.edu
Gelu's alma mater, where he now funds scholarships.
- Gelu Sulugiuc on LinkedInhttps://www.linkedin.com/in/gelusulugiuc
Connect with Gelu.
- Ryan Estes on LinkedInhttps://www.linkedin.com/in/estesryan/
Connect with the host.
- AI for Foundershttps://aiforfounders.co
The newsletter and podcast this episode comes from.
- Inbox Alchemyhttps://inboxalchemy.co/
Ryan's tool for building a newsletter audience you own instead of renting from a platform.
- CEOfriendhttps://bit.ly/3SrjgL1
Free competitor analysis from this episode's AI advisory board sponsor.
- AI Native Studenthttps://ainativestudent.com/
AI skills training for job seekers, this episode's sponsor.
- Momentoushttps://crrnt.app/MOME/8RDrnXDd
35% off any supplement subscription, this episode's sponsor.
- Taelorhttps://taelor.style/
An AI men's fashion stylist and this episode's sponsor; use code RYAN30 for $30 off.
- Get Howdihttps://gethowdi.com/
An autonomous AI sponsorship agent for podcasts, this episode's sponsor.
Named Frameworks
Your Model Is the Data
Why labeling matters more than architecture.
- →Model quality depends far more on labeling than on architecture.
- →Pair a domain expert (a journalist) with a machine learning expert to label documents correctly.
- →Each new model version mostly comes from improving the labels.
The Speed and Accuracy Double Lock
Being first only counts if you're also right.
- →Being first is worthless if the headline is wrong.
- →Track speed in milliseconds and the error rate every month.
- →Make both public. Gelu cites a record month with an error rate of about 0.13%.
Thesis, Budget, Trigger
Make the call calmly before the moment hits.
- →Do your homework before the event and write down your thesis.
- →Decide how much you're willing to lose on any single bet, then stick to it.
- →Pre-set the trigger so the reaction can be manual or fully automated.
Borrow a Giant's Distribution
Partner with incumbents instead of fighting for every seat.
- →Partner with an incumbent (Dow Jones Newswires) to reach about a million terminal users.
- →Add earned media (financial press, podcasts) as the second track.
- →Spend on product first and branding later.
The Secret-Edge Market
When customers won't share your product, grow another way.
- →When customers see your product as their advantage, referral programs stall.
- →The same instinct drives retention: paid users almost never cancel.
- →Grow through partners and press, not word of mouth.
Signal Over Sandpaper
Automate the watching so you only look when it matters.
- →Endless scrolling smooths out your brain one swipe at a time.
- →Define the information you actually need, then automate the watching.
- →Use alerts and watchlists so you only look when something matters.
Glossary
Latency
The delay between an event happening and the information reaching you. FNS measures it in milliseconds.
Earnings beat or miss
Whether reported results came in above or below what analysts expected.
Consensus estimate
The average analyst forecast for a metric. FNS compares results against FactSet consensus.
Structured data feed
News delivered as machine-readable fields (often JSON) so trading software can act on it automatically.
Algorithmic trading
Computer programs that place trades automatically based on pre-set rules.
Watchlist and alerts
A list of companies or keywords that trigger a notification when relevant news drops.
M&A
Mergers and acquisitions.
Share buyback
A company purchasing its own shares, often read as a signal of confidence.
Secondary offering or rights issue
A company raising new money by selling more shares.
13F filing
The quarterly U.S. filing showing institutional holdings. This is the “what hedge funds bought this quarter” data Gelu says FNS plans to cover.
Fundamentals
The full financial statements behind the headlines: income statement, balance sheet and cash flow statement.
Hallucination
When an AI model states something false with confidence.
Data labeling
Tagging documents so a model learns what matters in them.
Specialized model
A smaller AI model trained for one narrow task, usually faster and cheaper than a general LLM.
ICP
Ideal customer profile.
TAM
Total addressable market.
Churn
The rate at which customers cancel.
Pre-seed
The earliest institutional funding round, usually before or around product launch.
Bloomberg Terminal, LSEG Workspace, FactSet
Professional financial data platforms that cost thousands per seat.
23-hour trading
Exchange plans to extend trading to nearly the whole day.
