Not a chat widget. Voice and natural-language actions that create, update, and query real application state - with confirmations, guardrails, and the production infrastructure to run them safely.
Lenka Kadlec · Applied AI Product Engineer · Ex-Microsoft
Most teams are stuck at “we added a chat widget.” I move them to “users can actually do things through AI.”
The easy 20% - model calls, streaming, a chat box - is solved. The hard part is everything around it: getting an agent to trigger real state changes reliably, handling the failure modes, confirming before it acts, and keeping cost-per-action economically viable under real input.
That's the part I've actually shipped - and the full-stack depth to back it: auth, multi-tenancy, billing, and the infrastructure that keeps AI features from silently breaking in production.
Scoped, paid, and reassessed in stages. No open-ended retainers, no runaway scope.
Before any code, I assess whether your AI use case is viable - where an agent genuinely adds value versus where it's theatre - what the MVP needs, and what it'll cost per action. You'll know if it's worth building.
Architecture, stack, agent design, timeline, cost estimate, and risk areas - delivered as a written document you can take anywhere. Most clients stay.
4-week packages with clear deliverables and checkpoints. Both sides reassess and recommit at the end of each package.
AI that takes action is where I add the most value - backed by the full-stack engineering to ship it for real.
LLM tool calling and structured outputs that trigger real actions in your product - create, update, query, navigate - through natural language or voice. Human-in-the-loop confirmations, error handling, and guardrails so the agent acts safely, not unpredictably. This is the core of what I do.
Retrieval-augmented generation grounded in your data - with attention to chunking, source hierarchy, and keeping answers honest instead of hallucinated.
Agentic pipelines that replace manual steps: document processing, intelligent routing, multi-step automations with cost modelling built in.
End-to-end products: auth, multi-tenancy, billing, RBAC, real-time features - the production layer your AI features need underneath them.
Fast, accessible, production-ready web apps with App Router, RSC, and Tailwind. My primary stack, delivered to a high bar.
REST/GraphQL APIs on Node.js, Prisma, PostgreSQL. Redis caching, rate limiting, audit logging, OWASP practices, and performance tuning baked in.
Personal products and professional highlights.
A production-grade multi-tenant project-management platform with an AI assistant that triggers real in-app actions by voice - create projects, assign tasks, and query status through natural language, with confirmations and real state mutations rather than a chat wrapper. Real-time collaboration, role-based access, and subscription billing built in. Currently in beta, being hardened for release.
Shipped production features for Microsoft Teams messaging, used by millions worldwide. Led cross-team frontend delivery as primary engineer - coordinating rollout across multiple partner teams and working directly with PMs and principal engineers on scope, sequencing, and risk.
A full-stack app that turns any YouTube video into per-minute summaries, key points, and a “chat with this video” experience. Pulls the transcript, summarizes with Claude, extracts claims, and verifies each against web sources. Built end to end: durable background workflows, a multi-runtime architecture, streaming chat with prompt caching, and per-job cost tracking that surfaces real unit economics before you scale.
More work
A community recipe platform with AI-generated meal-plan population - users describe their preferences and an AI assistant fills their weekly plan. Built with Redis caching, Cloudflare edge delivery, S3 image storage, and a PWA shell for mobile. Full auth, rate limiting, and E2E test coverage.
Open to new projects, collaborations, and conversations. If you've got an AI feature stuck at the chat-widget stage - or a product to build from scratch - reach out.