PixelForge
Full-service web agency · 15-person team · 20+ active client projects
"We used to spend more time understanding bugs than fixing them. Now the AI gives us the root cause before we even look at the code. Our clients think we're wizards — we just have better tools."
PixelForge is a 15-person web agency based in Lyon, France, managing 20+ active client projects at any given time. Their client roster spans e-commerce, SaaS landing pages, corporate sites, and web applications. Each project has its own tech stack, its own stakeholders, and — until recently — its own chaotic feedback process.
The agency's core problem was translation loss. When a client reported a bug, it typically arrived as a Slack message: "The form on the contact page doesn't work." No screenshot, no browser info, no steps to reproduce. The developer assigned to the project would spend an average of 45 minutes per bug just trying to reproduce the issue — often failing because the bug was browser-specific, or dependent on a particular user flow the client didn't describe.
Multiply that by 30-40 bug reports per week across all projects, and PixelForge was burning over 25 developer-hours per week on reproduction alone. That's more than three full working days, every week, lost to context-gathering instead of building.
The team had tried various solutions. Some clients used Hotjar, others used BugHerd, a few sent annotated screenshots via email. There was no standardized process. Each client tool meant another login, another dashboard to check, another billing relationship to manage. For the smaller projects on tight budgets, there was no tool at all — just Slack messages and hope.
Client satisfaction was stagnating. Their NPS had plateaued at 42 — decent, but nowhere near the 70+ that top agencies target. Exit interviews revealed a consistent theme: "Bug fixes take too long" and "I feel like I'm repeating myself." Clients were frustrated that they had to explain the same bug three times across email, Slack, and a call before developers understood the issue.
PixelForge's Technical Director discovered FeedbackLoop AI's multi-tenant architecture and realized it solved their specific problem perfectly: one self-hosted instance, unlimited client projects, each with its own API key and configuration.
The deployment model was straightforward. PixelForge runs a single FeedbackLoop AI instance on their existing infrastructure. Each client project gets its own tenant with a unique API key, custom widget branding, and — critically — per-client language settings. Their French clients interact with the AI in French. Their English-speaking clients get English. The AI qualifies feedback, asks follow-up questions, and generates analysis in whatever language the tenant is configured for.
For each new client project, onboarding takes under 10 minutes: create a tenant in the admin panel, add the <script> tag to the client's site, and share the dashboard link. No per-client billing, no feature gates, no usage limits.
The transformation was immediate. When a client clicks the feedback widget and says "the form doesn't work," the AI asks clarifying questions — "Which form? What did you enter? What happened after you clicked submit?" — while simultaneously capturing the full session replay, console errors, network requests, and DOM state. By the time the developer opens the ticket, they have a complete picture: the exact form, the exact error (TypeError: Cannot read properties of null), the exact browser (Safari 17.2 on macOS), and a replay showing the user's entire interaction.
"The first week was eye-opening," Sophie recalls. "A client reported that their checkout was 'broken.' Before, that would have been two days of back-and-forth. With FeedbackLoop, we opened the ticket, watched the 30-second replay, saw the JavaScript error in the console panel, and pushed a fix in 40 minutes. The client got a Slack notification that same afternoon: 'Fixed.' They couldn't believe it."
After six months of running FeedbackLoop AI across all client projects, PixelForge measured the impact across their entire operation:
The agency also discovered an unexpected benefit: the frustration detection doubled as a quality assurance layer. During client staging reviews, FeedbackLoop AI's dead click and rage click detection caught UX issues the team's own QA process had missed. On one e-commerce project, it flagged a "Add to Cart" button that didn't respond on mobile Safari — a bug that would have gone live without the automated frustration tracking.
PixelForge now includes FeedbackLoop AI in every client proposal as a standard part of their service offering. Clients see it as a premium perk; the agency sees it as a cost-saving operational tool. Both sides win.
PixelForge's multi-tenant deployment uses a single instance serving all client projects. Here's how they configured it.
1. Multi-tenant configuration — each client gets a unique tenant:
Tenant: "Boulangerie Dupont" API Key: pk_bd_a1b2c3d4... Language: fr Domain: boulangerie-dupont.fr Tenant: "Nordic SaaS Co" API Key: pk_ns_e5f6g7h8... Language: en Domain: app.nordicsaas.io
2. Per-client widget embed — added to each client's site:
<!-- French client — AI speaks French --> <script src="https://feedback.pixelforge.dev/widget.js" data-api-key="pk_bd_a1b2c3d4..."></script> <!-- English client — AI speaks English --> <script src="https://feedback.pixelforge.dev/widget.js" data-api-key="pk_ns_e5f6g7h8..."></script>
3. Per-client language settings — the AI adapts automatically:
# French client sees: AI: "Merci pour ce retour. Pouvez-vous me dire sur quelle page cela s'est produit ?" # English client sees: AI: "Thanks for reporting this. Can you tell me which page this happened on?" # Both generate structured reports in their language # with session replay, console data, and root cause
The entire setup runs on a single Docker Compose stack. Client data is fully isolated by tenant. The agency dashboard shows all projects in one view, with per-client filtering. When a developer is assigned to a project, they see only that client's feedback, replays, and AI analysis.
Deploy one instance, serve unlimited clients. Per-client language, per-client branding, zero per-client cost.