Pricehound
Describe what you want to buy. An agent pipeline plans searches, reads reviews and forums, and returns a shortlist where every claim links to its source.
Hard $1/day AI spend cap, 1,460 testsThings I built, grouped by what they show. Employer and private code stays private, so for those I describe the work and give the numbers.
LLM products and agents. The interesting part is what stops them going wrong.
Describe what you want to buy. An agent pipeline plans searches, reads reviews and forums, and returns a shortlist where every claim links to its source.
Hard $1/day AI spend cap, 1,460 testsAn AI filter kept missing good deals. I built an eval set from 1,044 real listings and tested every prompt change against it.
Caught 23 of 23 known deals, up from 8Servers that let agents book appointments and run social pages. Each one only exposes what the agent may touch.
Nothing outside the allow list, every write loggedAn expenses app that reads bank statements with an LLM. If the balances don't add up, it refuses the file instead of guessing.
Account numbers stripped before every AI callFive years of high-traffic, transactional backends. Employer code stays private, so these are the numbers.
Moved a high-traffic transaction backend from REST to gRPC, with load tests before and after.
P99 latency down 59%A matchmaking service built with CQRS and a message bus, running in Docker on Kubernetes.
500 req/s with P99 under 100 msA starter for transactional .NET services: Clean Architecture with CQRS. Aspire and Pulsar are optional.
Clean Architecture and CQRS, ready to extendThe platform my own projects run on.
About 15 apps on one server behind a tunnel and SSO, with push-to-deploy and nightly encrypted backups.
No inbound ports open to the internetA web UI to run and watch coding-agent sessions on that server, including what each one costs.
About 950 tests, no frameworkLepri is my studio for local businesses. Each one has a short write-up and a live site.
A car import catalogue with an AI assistant that answers in Albanian, plus an admin CRM.
Live, with an AI assistant in AlbanianA sneaker shop and admin CRM. An order opens a pre-filled WhatsApp message instead of a checkout.
Live, orders over WhatsAppA fencing company's site with a quote request form.
Live, quote requests by formA landscaping company site the owners update without a developer.
LiveEmail is the fastest way to reach me.