About FASHN AI
FASHN AI builds the generative models behind virtual try-on, AI fashion models, and model-swapping for fashion brands, turning a single product photo into studio-quality campaign imagery. Founded independently by husband-and-wife team Dan and Aya Bochman, the self-funded startup runs on a long, fragile chain of AI endpoints: image analysis, captioning, generation, video, and uploading, each one a separate model, a separate provider, and a separate way to fail.
Debugging a pipeline like that meant solving two different problems. The first was better observability into their own application. They needed error tracking, tracing, and logs that were actually connected, so they could find and fix what was broken in their code. The second was keeping the product up and performant despite depending on AI endpoints they didn’t control.
“We’re using a lot of AI endpoints, and all of them have their own availability issues.”
— Dan Bochman, Co-Founder, FASHN AI
Sometimes the fix was a code change. Just as often, it was rerouting traffic to a different provider before customers noticed, and catching either one in time meant finding a tool built for both.
Outgrowing a tool built to do everything
FASHN had been using a general-purpose product analytics platform that had added so many products so quickly that none of them talked to each other. Technically, Bochman says, there was nothing wrong with it, it just wasn’t built around the thing he actually needed, which was understanding and fixing broken code. “I got lost in their dashboards,” he says. “Just missing a little bit of joy in the product.”
The team had Sentry integrated into their codebase for about a month when Bochman received Sentry’s weekly digest email.
“I got this email out of nowhere from Sentry, and I recognized the errors right away, I could clearly see what was triggering them and the impact on our users. This is something that only someone who really cares about developers would create.”
— Dan Bochman, Co-Founder, FASHN AI
Building that same signal into their existing analytics tool would have meant doing the engineering work Sentry had already done. With the confidence the team had in its error monitoring, they expanded into Sentry’s other offerings.
“If you have an error, you want to see the logs and spans around it. Either the thing that caused it, or just the context surrounding it, all in one place. In the age of AI, that context includes the agent conversation: if the error happened mid-conversation, you want to see exactly how it went, the prompts, and the tool calls. Sentry gives us all of that.”
— Dan Bochman, Co-Founder, FASHN AI
The challenge: improving uptime
AI is not reliable infrastructure. At any given time, something can be down, slow, overloaded, or rate-limited — and for FASHN, whose pipeline depends on a chain of foundation models, that meant someone had to make a judgment call every time an endpoint degraded.
“There’s this moment when your video generator endpoint goes down. Now it’s like a faith game: do you wait and hope it’s just a five-minute hiccup, or do you stop everything and manually redeploy to a backup, and then redeploy back once it’s over? And maybe you did all that for nothing, because it really was just a five-minute hiccup.”
— Dan Bochman, Co-Founder, FASHN AI
That guesswork was expensive either way. FASHN’s serverless workers used to make this call independently, hitting a rate limit, deciding an endpoint was “too hot,” sitting on a cooldown. There was no shared picture of what was happening across the fleet. Left unresolved, the cost showed up as wasted engineering time at best and extended outages at worst. “There were days we were in four hours of downtime because something that usually would have been solved in 30 minutes just didn’t solve itself,” Bochman says.
Rebuilding per-endpoint visibility
Fixing this meant two things: seeing exactly which part of the pipeline was actually failing, and then acting on that signal without a person in the loop.
FASHN had built the first version of that visibility on Fal.ai, whose SDK gave logs per endpoint and per request. When FASHN switched infrastructure, that visibility disappeared. With Sentry, they rebuilt it at a finer grain. Each step in a generation pipeline — analyzing, captioning, generating, uploading — is instrumented separately, so they can isolate which component is actually slow.
That’s how they traced a “generations are slow” complaint to a Cloudflare upload bottleneck that no single endpoint’s health check had caught. Only after ruling out their own code did they confirm the cause was on Cloudflare’s side.
“Finding something like this would have taken hours of digging through logs and code by hand. Now Sentry breaks it down by component, so we know exactly where to look.”
— Dan Bochman, Co-Founder, FASHN AI
Building a closed-loop rerouting workflow on Sentry
When an external API endpoint in one of FASHN’s agent workflows started showing latency problems, the team used Sentry to define what “healthy” looks like for that endpoint, then built an alert around deviation from it — a heuristic like requests running slower than five minutes for more than two minutes straight. Sentry fires a webhook the moment that threshold is crossed, and from there it’s hands-off:
“We get alerted via Slack, but also our server is alerted and changes things automatically. It’s hands-off. We are switching providers and falling back to different infrastructure automatically, completely with Sentry’s webhooks.”
— Dan Bochman, Co-Founder, FASHN AI
FASHN uses Sentry webhooks to automatically reroute traffic to fallback providers based on performance thresholds
FASHN still routes 10 to 20 percent of traffic to the original provider after rerouting, so Sentry can detect recovery and send traffic back automatically. They have built a closed loop in both directions, with no engineer paged and no manual redeploy. “With Sentry’s help we’ve built a closed-loop system that keeps uptime high and engineers from being paged,” Bochman says.
It’s built on Sentry’s alerting rules and open API rather than the UI. What sold Bochman was an API he could pull Sentry context from directly, the alert rules he could define himself, and webhooks that fire the moment those rules trip.
“This was one of the main reasons we chose Sentry. I wanted someone who really cares about performance and developer experience. I’m not just looking for UI, I want to access this data and for it to be actionable.”
— Dan Bochman, Co-Founder, FASHN AI
The team is now evaluating which other points in the pipeline deserve the same treatment.
“If something starts as a monthly hiccup and becomes a daily hiccup, we see it through Sentry and then decide to set up automatic rerouting, because the data is already there. With Sentry you have saved us the stress and mental energy of hours of debugging.”
— Dan Bochman, Co-Founder, FASHN AI
Beyond code: catching marketing failures too
The same alerting pattern has shown up in unexpected corners of the business. Bochman wired Sentry into FASHN’s Google and Meta ad conversion tracking after struggling to confirm data was reaching those platforms reliably.
Now an alert fires the moment an upload to those trackers fails. “It became a habit,” he says. “Otherwise I would have to go through the logs, or look at the Meta or Google dashboard to see if I can see conversions going through. But now I can rely on Sentry, get alerted, and know if data is flowing.”
Why it works for them
- One place to debug. Errors, logs, traces, and AI conversation context all live together and connect back to the same issue. FASHN no longer pieces together what happened by jumping between tools.
- Focused and detailed, instead of trying to do everything. FASHN moved off a product-analytics tool that covered a lot of ground but treated errors as one feature among many. A platform built around errors specifically helped developers save time and improved customer experience.
- Action, not just visibility. Programmatic access to alerts means FASHN can reroute traffic, flip feature flags, and revert automatically without a person in the loop.
- Granular, per-component tracing. Each step of a multi-stage AI pipeline — analysis, captioning, generation, upload — is instrumented with Sentry, so the team can isolate exactly where a slowdown lives.
- Faster triage. Issues that used to take about an hour to identify and fix now take 5 to 10 minutes.