Payment platforms, marketplaces, IoT analytics and AI systems
for startups and enterprises in fintech, healthcare, energy and automotive
A SaaS serving many companies had to take payments correctly when two customers check out at once, and keep each company's data sealed off. We hardened checkout against double processing, enforced tenant separation on the server and caught corrupted translations before users did.
An app where teams work with AI agents holds API keys, queued messages and settings - where a quiet bug does lasting damage. We fixed five defects that lost data or exposed secrets, from dropped messages to API keys leaking in streamed replies, each proven by a test first.
Payroll is where a small calculation error becomes a real payment. We built our own payroll system with exact part-month commissions, access that updates the moment a role changes, and business rules enforced on the server.
Custom Rust software that runs ONNX models on AMD GPUs through WSL/ROCm, splitting a local RAG workload across a discrete and an integrated GPU so search stays available while indexing runs.
An MCP platform with a dedicated test framework for tool quality and a bridge that lets local models on Ollama, ONNX and vLLM use sandboxed files, GraphRAG and code-RAG tools - up to 10x higher task effectiveness with models up to 32 GB.
An AI wellness logging and coaching app - chat-like meal, workout and weight logging with photo and barcode capture, and an AI coach with tool calling and memory - where a 500-case QA cycle raised the pass rate from 15% to 83.8% and a two-tier vision strategy cut image costs ~7×.
A code-intelligence platform combining LLM-generated semantic descriptions, Qdrant vector retrieval, PostgreSQL full-text search and Neo4j dependency graphs - so developers and AI agents can search large codebases by meaning, identifiers, call hierarchies and architecture.
An automated reply system for private messages and comments across Facebook, Instagram and TikTok, using retrieval-augmented generation for accurate, brand-consistent responses - with one abstraction over Claude, GPT and other providers for easy switching and cost control.
An AI analytics platform that scores real engineering effort per pull request - semantic task alignment, code constructs and change type rather than raw lines - tracks developer output over time and quantifies whether AI tooling actually improves delivery.
For a company with 100+ developers, an AI estimation and sprint-planning system - initialised on company data and refined after every completed task - reached ~90% agreement with the strongest developers' estimates after six months, cutting sprint planning to hours.
A client's four AI coding tools ran continuously against direct token APIs at unsustainable monthly cost. A redesigned execution layer with quota-aware routing across licensed subscription runtimes cut token cost roughly 1,000-fold - with no workflow change for developers.
A multi-agent development platform - Claude Code, Codex, Gemini and local LLMs orchestrated on .NET 10/Aspire - that takes Azure DevOps, Jira or free-text work items to specs, code, tests and pull requests under human approval gates. Proven on two commercial projects.
Tell us about it - we will show you how we approached comparable projects and what it would take.
