CurlyFry: a production AI SaaS, built end to end

What it is

Engineering managers live across five tools: meeting notes in one, delivery metrics in another, tickets, dashboards, and a spreadsheet holding it together. CurlyFry replaces that with one product: 1:1 intelligence, sprint command center, incident triage, knowledge-graph views of bus factor and mentorship, OKR tracking, and AI summaries threaded through all of it. React 19 and TypeScript in front, FastAPI and PostgreSQL behind, Celery workers on Redis for the background machinery.

The part most AI products get wrong: cost governance

Every LLM feature in the product routes through a gateway we built that picks the model tier by task type and customer plan. Summarizing a standup does not need the same model as analyzing a quarter of delivery data, and a $29 plan should not generate $60 of inference. The gateway enforces per-plan budget caps with graceful degradation down the model tiers, caches responses in Redis, falls back across providers when one has an outage, and supports per-organization bring-your-own-key with encrypted key storage. PII is scrubbed before anything leaves the building.

The error loop

The feature we're proudest of is invisible to users. When an unhandled exception occurs, the platform issues a short reference ID and stashes the full traceback and request context. An analyzer then pulls that context plus the surrounding pod logs from the log aggregator, hands the bundle to an LLM to classify the failure, and can open a draft merge request against the repository with a proposed fix.

Operating it like we tell clients to

• Kubernetes with GitOps: managed K8s, Helm chart, CI pipeline that lints, tests, builds, and deploys on merge. Horizontal autoscaling from 2 to 6 replicas.
• Multi-tenancy enforced by CI: every table is scoped to an organization, and a build-failing test verifies that any new foreign key is wired into cascade-safe account deletion. You cannot merge a table that would orphan customer data.
• Observability we own: log aggregation, dashboards, and health monitoring feeding alerts to chat. The same stack we deploy for clients.
• Two distributions, one codebase: the SaaS and a self-hosted enterprise edition ship from the same repository, with cryptographically signed license files and environment-driven feature flags. Live billing runs through Stripe.

The honest postscript

We built CurlyFry as a product bet, and the engineering outran the go-to-market. What it proves is the part we sell: one senior team taking an AI product from empty repository to operated, billable, multi-tenant production software in a quarter, with the cost controls and guardrails that make AI features economically survivable. If you want that built inside your company, that is the engagement.