Payments & fintech product guy: card payments, POS, open banking. Ex-PMO lead. Building AI agents and Brifco, an iPhone app that reads you your day. ๐ธ๐ฎ
Hi, I'm Feliks. About 6 years in card payments and fintech: issuing, acquiring, POS terminals, SoftPOS, card-processor migrations, open banking.
Now I build AI agents for ops work, and an iPhone app called Brifco.
What I'll post about here ๐งต
@deepakjain_in No loss history usually means conservative and staged: a bigger reserve or prefunding up front, tight per-card limits and MCC allowlists, then the reserve steps down as real loss data builds up. For agents I'd also want a human approval step above a set amount.
Two launches this month. One was announced in 2023 and delayed twice. The other was built by a small team in Slovenia and shipped on time. ๐ ๐ฎ GTA VI: out today ๐ฑ Brifco: your weather, calendar, traffic and news as a short audio show ๐ Sound on. Link below. ๐ #GTA6
@nat3hall@bot love the fake commercials. my guess is the bits are fun for a week, then you want the calendar up top and the jokes after. my cofounder and I built brifco, an iphone app that does the short one-host version of this. do you still listen to the whole show?
@TreasureMusaAI even "declined" isn't one state. insufficient funds or issuer unavailable can be retried later, a lost/stolen card or closed account never should be. retry logic that treats them the same racks up network fees and can get the merchant flagged
@Fahadh_4444 and don't put everything in one shared queue. give each customer's burst its own lane with a cap, so your own urgent changes never wait behind someone else's 10k creates. same pattern works for outbound webhooks and payment API calls
@elie2222 that email is the hidden cost of billing migrations. anyone moving now: email customers yourself before cutover, and make sure the new statement descriptor is recognizable. a confused subscriber who sees an unknown name on their card statement tends to dispute it
@vigstar1 looking forward to it. my bet for the slowest side is issuers: an agent-initiated purchase looks nothing like the sessions their fraud models were trained on, so they'll decline conservatively until it's clear who eats the dispute when the agent buys the wrong thing
@gaspardlezin part of it: inside the eurozone, SEPA Instant already gives people 10-second, 24/7 transfers, and the EU instant payments rules cap the price at what a normal transfer costs. euro stablecoins have to win on cross-border and onchain use, not domestic payments
@rodri_shema one more generated code gets wrong a lot: marking an order paid when the user lands on /success. users close tabs and redirects get replayed. only the signed webhook from the processor should flip it to paid
@aftaab___ that sucks. when comparing replacements, ask two things before fees: can they import your active subscriptions with the saved cards (that's a processor-to-processor PCI data transfer, not a CSV), and are they merchant of record so global VAT/sales tax stays their job
Card payments hit this every day.
A terminal sends an authorization. The issuer approves, but the response never makes it back.
The terminal doesn't retry. It reverses the original request, so the customer isn't stuck with a hold for a sale that never finished.
Imagine an application sends a database commit request.
The database commits successfully, but the network connection drops before the confirmation reaches the application.
From the application's perspective, the result is uncertain.
Blindly retrying the operation can create duplicate records or repeat business actions.
This is one reason reliable systems need idempotency, transaction identifiers, and a way to reconcile uncertain outcomes.
Distributed failures are difficult because sometimes neither side has enough information to know what happened.
@DanielSmidstrup Feliks from Slovenia. ~6 years in card payments and fintech product, now building AI agents that do real ops work, plus Brifco, an iPhone app that turns your day into a short audio show. respect for sharing real numbers, that's rare here
@Josephyala in payments the fix starts with admitting a third state: unknown. don't show the user success or failure, mark it pending, look it up by your own reference, and let the next settlement/recon file be the final word. double charges love systems that force unknown into yes/no
@oost_marcel@dLocalPayments the gap I'd watch: volume +77% but gross profit +28%, so each processed dollar earns less. normal when big existing merchants scale and get better pricing. 140% NRR plus falling opex turns that into profit, as long as those merchants don't start routing to a second provider
@krunalbuilds how many schedulers are running it now. an extra instance after autoscaling or a failover node that also fires cron is the classic one. then clock changes and lock leases expiring mid-run. and if that job moves money, it should be safe to run twice anyway
@balance0014 past demos: Brifco is live on the App Store for iPhone. it's a consumer app though, the ops agents are a separate thing I build for businesses (invoices, inquiries, booking). post-quantum is a big one for payments down the line
Sunday #connect for people on the money side of tech.
Payments, fintech product, card issuing/acquiring, open banking, or AI agents doing real ops work.
I'm Feliks: ~6 years in card payments and fintech, now building AI agents and Brifco.
What are you working on?
@SpecialeRob the "correct it twice, it becomes a rule" part is the bit I'd steal. one add: tag each rule with the date and the case that created it, so when a rule starts blocking good work you can check if that original exception still exists
@triston_armstr@NetworkChuck a morning report with industry news baked in is the right shape imo. what helped us: keep news to 2-3 stories and say the source first, otherwise it turns into a feed you skim. we're building that as an iphone app, brifco. how long does yours run?
@LiveLoveSEO two follow-ups to ask right after: is the signature checked against the raw body (parsed JSON breaks it), and what happens when the same event arrives twice? webhooks get retried, so store the event id and skip ones you've already handled