Prototypes are easy.We build the one that runs on Monday.
Nebuchadnezzar Network is an AI and data engineering studio. We join teams who have proven an idea and now need it to survive real data, real users, and a finance review — pipelines, retrieval, evaluation, and the unglamorous operational work underneath.
Three things we were brought in to fix
Each of these started as a prototype somebody was proud of and a problem nobody could schedule around.
Claims retrieval
Forty years of scanned claim documents sat behind a keyword search that adjusters had learned not to trust.
Lookup time fell from 14 minutes to 40 seconds across 1.2M documents, with citations back to the source page.
Nothing could leave the customer’s own tenancy.
Evaluation harness
A routing assistant shipped on screenshots and opinion. Every model upgrade turned into a two-week argument.
A 900-case golden set now gates deploys. Model changes land in a day and regressions are caught before customers see them.
Pipeline rebuild
Nightly reconciliation ran for nine hours and failed often enough that someone was paid to watch it.
Rebuilt as incremental jobs with explicit data contracts: 35 minutes, with failures isolated to one partition.
Migrated in place, with no reporting downtime.
Things we have shipped
Products and platforms running in the open. Some we built end to end, some we joined partway and helped finish.
Flight search, reinvented with AI for experienced and more luxury travelers
Jobbigt.nu is a Swedish / Norwegian job board search engine consisting of a shared backend that serves multiple boards
Fan site for the Eurovision Song Contest packed with statistics and data from every single Eurovision event taken place, since 1956.
Hydra is a modern CMDB with automation and CI/CD workflows in its heart. Launching in 2027
This site. A static Astro build with schema-validated content, a browser editor that commits to GitLab, and atomic deploys to our own server.
How an engagement runs
Four stages, each one consuming the last. You can stop after the first and still have something worth the money.
Map
2 weeks
We read the code, the data, and the last six months of tickets, then talk to the people who get paged. Fixed price, and it ends whether or not we work together after it.
You receive: A written assessment and a costed plan
Build
6–12 weeks
We work in your repository, on your board, in weekly increments you can see running. No separate vendor environment that has to be integrated later.
You receive: Working software in your production account
Prove
Ongoing
Evaluation sets, load tests, and cost traces run before launch and keep running after it. If we cannot measure the improvement, we treat it as not having happened.
You receive: An eval suite and a cost-per-request figure
Hand over
2 weeks
Pairing, runbooks, and decision records until your team ships a change to the system without asking us. Then we leave and stay reachable.
You receive: Documentation, and an engineer who owns it
What we actually do all day
Most engagements touch four or five of these. We would rather do a narrow thing properly than list capabilities we would have to learn on your budget.
- Data platforms
Pipelines, storage, and the contracts between them. Batch or streaming, with lineage you can audit when a number looks wrong.
dbtAirflowPostgresIceberg
- Retrieval and search
Ingestion, chunking, embeddings, hybrid search, and rerankers tuned against your corpus rather than a public benchmark.
pgvectorOpenSearchrerankers
- LLM applications
Tool use, orchestration, and guardrails. Built on the Claude API or whichever model actually fits the latency and cost you have.
Claude APIMCPagent loops
- Evaluation
Golden sets, offline scoring, and online checks, so changing a prompt or a model is a decision you can measure instead of argue about.
eval harnessesLLM judgesregression gates
- Production and cost
Caching, batching, routing, and tracing. What a request costs today, where that money goes, and how to make the number smaller.
OpenTelemetryprompt cachingbatch APIs
- Handover
Runbooks, decision records, and pairing with your engineers until they own it without us. Leaving cleanly is part of the work.
ADRsrunbookspairing
Tell us what is stuck
One email, no form. The more specific you are, the more useful the first reply will be — we would rather read three awkward paragraphs than a calendar invitation.
- What exists today, and what it gets wrong
- Who it is for, and how many of them
- What has to be true for this to count as finished
We answer every enquiry within two working days.