The expensive bugsdon’t crash.

1. 01Personal capstone2026

   Credit-card onboarding pipeline→

   A duplicate message here is not wasted work. It is a second hard enquiry on a real person's credit file.

   • Java 17
   • Spring Boot 3.5
   • PostgreSQL 15
   • RabbitMQ
   • Resilience4j
   • Flyway
   • ·7 ADRs, including the ones that record what is not built yet
   • ·138 tests; one runs the whole pipeline against real Postgres and RabbitMQ
   • ·8 Prometheus alert rules, 4 Grafana dashboards
2. 02Own open source · MIT2026

   splitlint→

   The broken chunk still joins back into the original string. Nothing throws; retrieval just gets worse for one language.

   • Python
   • Unicode UAX
   • pytest
   • GitHub Actions
   • ·766 of 766 official Unicode grapheme conformance cases pass
   • ·LangChain's RecursiveCharacterTextSplitter fails 10 of its 12 cases
   • ·Grew out of an upstream fix in langchain4j
3. 03Open-source contribution2026

   AST-RAG: code retrieval for AI agents→

   Text chunks cannot answer "who calls this?". The graph can, but only if every file is parsed correctly and the parser is safe to share between threads.

   • Python
   • Tree-sitter
   • Neo4j
   • Qdrant
   • SQLite
   • ·Most of the repository's commit history is mine
   • ·Parse cache: content-addressed, bounded, two backends
   • ·Fixed a data race that corrupted results instead of crashing
4. 04Own open source · MIT2026

   Plumb→

   Ask an agent whether the feature works and it will say yes, sincerely, having verified nothing. "Done" was never defined, so the only available answer is a feeling.

   • Python 3.10+
   • Claude Code hooks & skills
   • mypy --strict
   • GitHub Actions
   • ·Zero runtime dependencies, no network, no telemetry
   • ·Uninstall removes only files whose hash it recorded
   • ·Never changes the user's permission mode
5. 05Professional · a global bank2023–2026

   Legacy service modernization→

   Behaviour had to be provably identical before and after, including the parts nobody had documented, because the thing on the other end of the endpoint is somebody's account.

   • Java
   • Spring Boot
   • PostgreSQL
   • REST
   • SOAP
   • Elasticsearch
   • ·17+ endpoints migrated, ≈35% of the programme
   • ·Test coverage raised to 80% by writing the test first
   • ·Legacy maintenance overhead down ≈25%

The onboarding service, simulated from its real configuration: retry counts, backoff, breaker thresholds, the outbox poll interval. Pick a failure. Every path ends in a named, audited state, except the one marked as a gap, which is written up with its fix.

Most of my AI work is on the unglamorous side of retrieval: getting text in without losing it, cutting it without corrupting it, indexing structure, measuring whether the right passage comes back, and making agents prove their work. Each stage below links the evidence.

1. 01Extract→

   Turn HTML, PDFs and office files into text.

   Breaks:Content is dropped or garbled, and the document still parses.

   • trafilatura #940mergedText directly after a list, table, quote or code block was dropped from every output format.
   • unstructured #4502mergedDefinition lists were discarded from HTML, so a Sphinx API reference page partitioned to its headings and nothing else.
   • unstructured #4449mergedspaCy ran three times on the same text per element: 88% of partition_html() time in the profile. Now once.
   • docling #3985mergedCSV conversion failed when the first quoted field spanned a newline: the dialect was sniffed from a truncated line.
   • docling #4388reviewMHTML pages with a meta charset were decoded twice, turning non-ASCII text into mojibake.

2. 02Split→

   Cut text into chunks small enough to embed.

   Breaks:A cut lands inside a character. The chunk round-trips and embeds as garbage.

   • splitlintAudits any splitter for it
   • langchain4j #6091mergedThe sentence splitter returned nothing for text with no detectable sentence, so the hierarchical splitter dropped it.
   • langchain4j #6308reviewSplitting by UTF-16 code unit cut surrogate pairs in half, producing chunks that cannot be encoded as UTF-8.

3. 03Index→

   Represent content so it can be found: vectors, and structure where it exists.

   Breaks:Structure is flattened away, or a race corrupts the index without failing.

   • ragedASTs, call graph and vectors for code

4. 04Retrieve→

   Find the passages that answer the question.

   Breaks:A confident answer from the wrong source, and nobody measured how often.

   • Ask this siteBM25 + embeddings + rank fusion, with an eval set

5. 05Act

   Agents call tools and change things.

   Breaks:The agent reports done without evidence, or the tool wiring quietly points elsewhere.

   • PlumbMakes coding agents prove they are done
   • mastra #20591mergedNine storage adapters declared @mastra/core peer floors older than the exports they import; installs passed, then died on import.
   • chainlit #3013reviewPer-user OAuth for HTTP MCP server connections.
   • hermes-workspace #735reviewThe MCP server list was fetched from the wrong endpoint on both the probe and data paths.

Fixes sent to libraries I use. Each one arrives with a regression test that fails before the patch and passes after it, because that is the only “I fixed it” a maintainer can check in a minute. The list is generated from the GitHub API; closed-unmerged PRs are left out.