Sveinbjörn Geirsson

# Engineering work

Three accounts of my commercial engineering work, with the contribution and limits stated explicitly.

## Search decisions backed by relevance tests

Massif Network · August 2025–present

I built hybrid text and image search for a location and media workspace used by film-production teams, using Meilisearch and Voyage AI embeddings. The engineering question was whether a change improved the returned locations, rather than merely adding more searchable data.

I built a graded golden set of 87 queries and an evaluation harness. I compared one vector per photo with one combined text-and-image vector per location, and tested the balance between keyword and vector search. The combined representation performed better overall in that comparison.

A separate 670-case A/B test added geographic information from Overture Maps to embeddings. It did not detect a relevance gain (p = 0.75). I recorded the result and recommended against paying to re-embed on that evidence.

These are two different experiments. The result supports a decision for the tested corpus and judgments; it does not prove that geographic enrichment never helps, or measure user satisfaction, revenue, or agent-task success. The underlying evaluation reports are private; this account summarizes my work rather than offering independently reproducible public data.

## Background work with visible failures

Massif Network · August 2025–present

Background tasks originate in a PostgreSQL queue (PGMQ) and pass through a broker Worker to domain Workers. I built this processing path using Cloudflare services including Hyperdrive, service bindings, and Durable Objects.

Processing uses time-limited claims, bounded execution, and retries. Claim fencing rejects stale acknowledgments and finalization of queue work. Tasks that exhaust their retries reach a dead-letter queue, giving failed work a visible place to investigate.

This makes the failure path explicit: a failed task can be identified instead of disappearing into an opaque workflow. Grafana Cloud metrics, logs, dashboards, and alerts provide additional operational visibility.

This is an implementation account, not a measured uptime improvement. Queue fencing is not a claim of exactly-once execution, protection against every downstream side effect, or a universal no-loss guarantee. The commercial source code is private.

## Settlement, reconciliation, and closing the accounting day

Korta / Rapyd · 2013–2022

I joined Korta directly from university and worked on merchant settlement, card-network reconciliation, and financial data services. The work required translating payment and accounting rules into software that other teams could use.

I developed fee, reserve, refund, splitting, withholding, and payment-frequency logic for domestic and SEPA/SWIFT transfers. I mapped Visa/Mastercard reconciliation and chargeback data to accounting codes, currencies, and cost/value dates, and implemented conditions for closing an accounting day.

I added logging and traceability to complex financial processing and automated releases. IBM DB2 data services made financial data available to frontend teams without requiring them to understand the underlying systems.

This account is supported by my recollection and earlier CV descriptions, not a public code audit. It does not claim transaction volumes, a measured reduction in failures, or sole ownership of the payment platform.

## Public contributions

Separate public work, linked here for direct inspection.

- [github.com/can1357/oh-my-pi/pull/789](https://github.com/can1357/oh-my-pi/pull/789)

- [github.com/tschk/crepuscularity/pull/299](https://github.com/tschk/crepuscularity/pull/299)

- [www.npmjs.com/package/pdq-wasm](https://www.npmjs.com/package/pdq-wasm)

- [github.com/Raudbjorn/ggml-llama.cpp/pull/35](https://github.com/Raudbjorn/ggml-llama.cpp/pull/35)