Backends, APIs and AI systems built to be operated — typed contracts, real observability, and failure modes decided in advance.
Most backends fail slowly. They work at launch, accumulate special cases, and become the thing nobody wants to touch. We build against that: explicit contracts, boring data models, and instrumentation from the first commit rather than the first outage.
REST and typed RPC surfaces with versioning, pagination, idempotency and error semantics settled before the first endpoint ships.
Retrieval-augmented generation over your own documents — chunking, embeddings, vector search and evaluation, built so answers stay traceable to a source.
Relational schemas that hold up under change, migration discipline, and caching layered in where measurement says it belongs.
Structured logging, tracing, health checks and alerting tied to user-visible symptoms rather than CPU graphs nobody reads.
Projects we shipped in software engineering. Every one links to running software or the source it was built from — check any of them.
2026Java
Desktop knowledge assistant answering questions over private document collections.
Full RAG pipeline in Java — document ingestion across TXT, PDF, DOCX and code files, embedding and retrieval through LangChain4j, and a JavaFX 23 client with full-text search and export.
2026Next.js
AI email client that drafts replies matching the sender's own writing style.
Next.js application with a three-pane client, multi-account state on Zustand, full-text search, and a tone-controlled drafting layer over an LLM.
2025TypeScript
University career services portal covering job matching, resumes and advisor scheduling.
Designed the matching model and the application workflow behind student dashboards, one-click apply and an events calendar.
2026LangChain
Retrieval endpoint answering questions from a live content corpus.
Vector store on DataStax Astra DB, retrieval orchestration through LangChain, Upstash Redis rate limiting and streaming responses over the Vercel AI SDK.
2026Python
Open-source Python classifier that found an audience on GitHub.
Detection pipeline and CLI, maintained in the open — the project has drawn stars and outside contributions since release.
The open-source projects this work stands on. Built and maintained by their own teams, not by us — listed here because knowing what we build with tells you more than a logo wall would.
Domain and data model first. Almost every expensive backend problem is a modelling mistake that shipped.
The API surface gets written and reviewed before implementation, so clients can build against it in parallel.
Typed end to end, tested at the boundaries, deployed continuously behind feature flags.
Dashboards, alerts and a runbook handed over — so the team inherits a system they can actually run.
What clients ask before starting software engineering work with us.
Web development is what the user touches. Software engineering is the system behind it — data models, API contracts, background processing, integrations, AI pipelines and the operational tooling that keeps it running. Plenty of projects need both; they're separated here because they fail for different reasons and get scoped differently.
Yes, and it's most of what this work looks like now. Retrieval over your own documents, drafting and summarisation, classification and routing, natural-language search. We start by asking whether the feature needs a model at all — a good number of ‘AI’ requests are better served by search and a decent data model, and we'll say so.
Retrieval-augmented generation grounded in your own content, with every answer traceable back to the source passage it came from. We build an evaluation set during the project so quality is measured rather than vibed, and we constrain the model to say it doesn't know instead of inventing a confident answer.
At the boundaries — API contracts, data access, and the logic where a bug is expensive. We don't chase a coverage number. Tests that cost more to maintain than the bugs they catch are a net loss, and we'd rather spend that budget on observability.
You get the code, the infrastructure definitions, the dashboards and a runbook. Support arrangements are available if you want them, but the handover is built so you aren't dependent on one.
Tell us what you're building. We'll come back with a read on scope, timeline and whether software engineering is genuinely the right call.