We used to write engineering blogs, covering everything from ML bias to cloud migrations. Then we went quiet to focus on building features.
I am restarting it with my engineering team. We will share technical challenges, code snippets, and details about our work.
Balancing fraud detection is always a trade-off. Be too aggressive, and you block legitimate customers and lose revenue. Be too lenient, and chargebacks cost you money. The right balance depends entirely on the business, whether you sell luxury goods or run a fast-food chain.
Here is what we shipped recently:
- Model Orchestrator: We rebuilt our orchestrator layer to route scoring requests to the right model ensembles, without needing separate deployments for every customer.
- Segment-Level Calibration: This catches model drift on specific events before it gets to workflows, ensuring seamless model releases.
- Sift Copilot: This was a winning hackathon project, an AI assistant navigates the Sift Console, explains workflows, and answers documentation questions.
- Claude Code: We adopted Claude Code across the whole team, and we will share metrics and lessons learned.
We will publish deep dives on:
- How the model orchestrator picks the right models per customer
- Segment-level calibration details and future plans
- Taking Sift Copilot from a hackathon project to a customer tool
- Cohort models and ensemble tuning
- Our Claude Code rollout metrics and lessons learned
After taking a pause, we’re ready to re-engage. Our engineering team is excited to share their work and insights.
We are hiring across platform, ML, and product engineering. Check out our open roles.
Ankush Walhekar, Sift Engineering
Estimated reading time: 2 minutes
