Data integration and reconciliation, including data written by automation and AI: I find why it quietly goes wrong. Nineteen years of backend and data systems. I work on data in motion — parsing and normalising what arrives from other systems, moving it, reconciling what does not match, and the analytics built on top. Java, Python, Go, SQL.
Most of what I publish is about the quiet failures — duplicates, gaps, stale values, order and reprocessing. They produce no errors in the log and surface weeks later as a wrong number. Automation and AI now write into accounting systems too — invoices read from PDFs, documents recognised from scans — and they make the same mistakes, only faster.
Reading unfamiliar code until the mechanism is named, with line references and, where possible, a test that shows it.
nats-server#8661 — when
max_ack_pending was not set, a configuration reload left open MQTT sessions with
a limit of zero, and QoS 1 and 2 delivery stopped. The fix and a regression test
were merged in #8671 less than
two hours after it was opened.
numaflow#3645 — an accumulator watermark that never advances when a key keeps receiving data and the function emits nothing for it. Verified with a local test against the Rust core. The team revisited the design, kept the behaviour as intended and restored the drop API in four SDKs; the documentation fix is merged in #3660.
nats-server#8687 — a stalled stream restore returned before the restore had stopped and without the completion advisory other failed restores send. The change and a test are approved by a maintainer in #8691 and wait to be merged.
nats-server#8607 — why
adding a stream source scans the whole stream, what opt_start_time actually
applies to, and what does bound the scan. Checked against the reporter's own
version rather than main.
ccxt#26773 — a balance update
lost rather than delayed, because deepExtend builds a new object.
Lean#9790 — three of
eight Send calls run off the result thread.
OCA/queue#996 — an Odoo job
recorded under the __name__ of the function it found, not the name it was
asked for. When a module installs a replacement under another name, as auditlog
does for write and unlink, the job fails the moment it runs. The name was
lost in four places; the fix and three tests are in
#998.
OCA/edi#1423 —
Odoo's PDF invoice import dropped a total printed with one decimal and turned an
integer into hundredths, without an error. Checked by running the conversion
against the module's own test cases and sample invoice: the proposed fix changes
none of them, while making the decimals optional would let a fragment of the VAT
number beat any smaller total under the max rule.
A watermark that wouldn't move — the numaflow case above, step by step: following a state clean-up condition through the code and checking it with a test.
What breaks in price feeds — 1,163 commits matching six feed-failure terms across 48 crypto organisations, and how the vocabulary differs between publishing a feed and consuming one. A hand check of 100 diffs found about one in six is an actual repair; the correction and every classification are in the repository.
What DAOs actually fund —
93 funding proposals across 29 governance forums: what gets funded, for how much,
and through which channel. analyse.py reproduces the dataset byte for byte.
As an independent contractor I worked directly for telecom companies in Slovenia — Iskratel and RC IKT — remotely, with acceptance testing on the client's own equipment.
crypto-scout — event-driven services that ingest market and on-chain events: collector, queue, analyst, TimescaleDB.
metacfg4j — a configuration library with a business abstraction over CRUD services, a DSL and an MVP.
I write the method down next to the result. If a number is published here, the code that produced it and the data it ran on are published with it — a disagreement should be with a rule you can read, not a figure you have to believe.
Limits go in before conclusions, not in a footnote.