Senior backend engineer in Valencia, Spain. I build PHP/Laravel services: APIs, queues and background jobs, SSO integrations, and the systems that keep releases running reliably.

In open source, I look for bugs in the tools I use, from AI SDKs and agent memory to queues and locks. Each report comes with a reproduction, a failing test and, where I can, a fix.

Merged so far: 27 fixes in Symfony, 9 in openai-php/client, 8 in symfony/ai, 1 in the official MCP SDK for PHP, 2 in Predis, 5 in Spatie packages. The Hindsight maintainer fixed all 5 of my bug reports, one with me as co-author.

Under review: over 50 more pull requests, in symfony/ai, Prism, both MCP SDKs for PHP, openai-php, LLPhant, the Gemini PHP client, Monolog and php-amqplib. What each one fixes.

My projects:

• hindsight-php: long-term memory for Laravel AI SDK agents, Laravel MCP servers, Symfony AI agents (v0.8, v0.9, with a Symfony bundle since v0.11) and Neuron AI agents (v0.10), on the open-source Hindsight server. Listed in the Hindsight integrations hub by its maintainers. v0.4 added recall, retain and reflect tools for laravel/mcp, so Claude Desktop or Cursor gets memory with a bank per user; v0.5 to v0.7 add the pages API of Hindsight 1.2.1, with tools to search and read pages for MCP clients and agents. The tools also work in Prism. Memories go through the Laravel queue, and a retry never stores one twice; Hindsight::fake() makes it testable. CI runs against real Hindsight 1.1.0, 1.2.0 and 1.2.1 servers.

• Symfony: twenty-seven merged bug fixes in Messenger, Lock, Semaphore, RateLimiter, Scheduler, Cache, HttpClient and the new JsonSchema component, each with a failing test and a reproduction:
  • #68021 RateLimiter: the sliding window counted reservations for later windows as hits that had already happened, so workers waiting for a reservation went over the limit: with 10 per minute, 50 workers got 60.
  • #67993 Messenger: with --fetch-size, --keepalive covered only the message being handled, so the rest of the batch went to another worker and was handled twice (reproduced with two workers on PostgreSQL).
  • #68020 and #68022 Redis transport: after a failed read or reject, a worker with the new cooldown strategy kept running but never read again the messages Redis had already delivered to it.
  • #68011 Semaphore: once a slot of a LockStore semaphore had expired and another process took it, acquire() still returned true, so two processes held a semaphore limited to one.
  • #67994 Lock: once the PostgreSQL session was lost, DoctrineDbalPostgreSqlStore never gave that lock to the process again until it restarted.
  • #68029 Lock: with a Predis client created from a DSN, Redis errors were taken for a success, so after a failover two processes could hold the same lock.
  • #68025 Scheduler: after a restart, a cron task without a timezone moved to the timezone of the task that ran last, so a nightly 02:00 UTC job ran at 07:00.
  • #68019 and #68088 RateLimiter: the token bucket counted the wait and the reset time from now instead of from its last refill, so retryAfter was up to a full interval late. The maintainer asked for the second fix after merging the first.
  • #68095 HttpClient: a retried PUT or POST with a stream or generator body went out empty, and the server accepted the request without its payload. Nicolas Grekas extended the fix to redirects and to streams nested in the body.
  • #68133, #68134 and #68135 JsonSchema, the new 8.2 component that describes tool inputs and structured outputs to AI agents in Symfony AI: with extra attributes not allowed, the objects created by a #[SerializedPath] stayed open, so OpenAI strict mode refused the schema; #[Count] and #[NotBlank] on a map such as array<string, int> gave it minItems, so {} passed; a default of INF or NAN made the whole schema fail to encode. Nicolas Grekas extended each fix before merging.
  • #68130 HttpClient: when a streamed POST lost its connection, EventSourceHttpClient sent it again without telling the caller, so a Symfony AI chat completion was generated and billed twice. Merged into 7.4.
  • Also merged: CompoundLimiter accepting a hit that one of its limiters refused (#68016), a Redis keepalive check comparing milliseconds with seconds (#68017), MongoDB isAcquired() failing before the first acquire() (#68027), Memcached locks longer than 30 days expiring at once (#68028), a Scheduler worker crashing when its lock expired during a long run (#68030), 'P1M' intervals running every 31 days (#68026), tagged cache items without expiry that were read back as expiring in 2090 (#68090), a task added to a running Scheduler that ran all its earlier slots at once (#68089), ChainAdapter losing the expiry of items from FilesystemAdapter (#68091), CachingHttpClient::stream() yielding a stream instead of its chunks (#68093), CachingHttpClient taking invalid HTTP dates for valid ones, so Expires: 3600 was fresh until the year 3600 (#68094), and the SSE parser giving an empty event: the type '' instead of message and dropping an empty first data: line (#68131).
• Predis (the PHP Redis client, ~7.8k stars), two fixes merged by Till Krüss:
  • #1756 once a transaction had sent WATCH or MULTI, a lost connection made Predis retry the command on a new connection, outside the transaction. With Sentinel this happened with default options, and a check-and-set transaction then overwrote a change another client made after WATCH.
  • #1755 when Redis rejected a command queued in a transaction built with the fluent interface (OOM, READONLY after a failover), the connection stayed in MULTI, so every later command on it was queued and never ran. With persistent connections, the next request on that PHP-FPM worker got the same socket.
• Spatie (Laravel packages), five fixes merged by Freek Van der Herten:
  • laravel-query-builder#1076 (~4.5k stars) ->ignore(null), the example in its docs, dropped filter[active]=false, so the query returned every record.
  • laravel-sluggable#320 (~1.6k stars) with self-healing URLs, a route that named its column, such as {post:id}, returned 404.
  • laravel-activitylog#1481 (~5.9k stars) with buffering on, an activity model without an updated_at column lost every buffered activity of the request.
  • laravel-settings#371 (~1.5k stars) settings with a DateTimeInterface or CarbonInterface property failed to load.
  • laravel-markdown-response#15, which serves pages as markdown to AI agents: with the middleware registered globally, as its docs suggest, /about.md returned the HTML page. Released in 1.2.4.
• laravel-queue-lab: reproducible experiments with Laravel workers, rerun in CI every week: jobs that run twice after RabbitMQ's acknowledgement timeout, a frozen broker that queue:work doesn't notice, rabbitmq:consume ignoring SIGTERM in maintenance mode, two Laravel 13 bugs: debounce maxWait that didn't force a single run in 60 s of steady dispatches, and a timeout counted twice against maxExceptions (v0.5), and a Laravel AI SDK agent in a queued job whose tool runs twice when its provider fails over after the tool call (v0.6).
• Queue reliability checks for other people's issues: an analysis of #664 (the reported hang is really redelivery), an independent check of PR #674, and measurements for php-amqplib#1227.

• Queues and workers: retries, connection failures, recovery, and safe restarts.
• Release reliability: what happens when old and new application versions run together.
• AI agents: MCP servers and clients, tool calling, and long-term memory that survives retries.
• AI-assisted development: understanding, testing, and reviewing the code that coding tools produce.

PHP · Laravel · PostgreSQL · Redis · RabbitMQ · Docker · Kubernetes

Connect on LinkedIn