Felipe Fumero

Creating software to learn how things actually work.

[Felipe Fumero @FelipeFumero663]

Currently building retrieval-quality-benchmark: a measured comparison of naive vector search vs. hybrid search + reranking, no hardcoded scores.

Portfolio → https://t.co/r1kMla1wcd
GitHub → https://t.co/0NLQRcOtCL
LinkedIn → https://t.co/uXp4s6VnNg
New angle posted daily.

[Felipe Fumero @FelipeFumero663]

@NoahSimons2012 A match-3 game with cloud saves and a global leaderboard is a lot more infrastructure than the genre usually needs, most match-3 clones skip the backend entirely. What's handling the leaderboard, something like Firebase, or did you roll your own backend for it?

[Felipe Fumero @FelipeFumero663]

@snapcallin Shipping the first version even with payments and a few features still missing is exactly the right instinct, most people wait until everything's done and never ship at all. What made you decide to cut it there instead of waiting on payments first?

[Felipe Fumero @FelipeFumero663]

@soni_sahil_0 A telehealth flow with OTP login, doctor tips, payment, and prescription meds all wired together in about a month is a solid freelance scope for that timeline. What ended up being the most annoying integration, the payment gateway, or the prescription/pharmacy piece?

[Felipe Fumero @FelipeFumero663]

@Smart_onX @apkpure Building a tool for the exact moment you stare at the compose box trying to think of something good is such a relatable problem to actually solve. Is the tone selection driving a different prompt template per tone, or one prompt with tone as a parameter?

[Felipe Fumero @FelipeFumero663]

@TheMaverick_tw Going from blank page to a polished animated landing page and actually shipping it, instead of endlessly polishing, is the harder discipline. What ended up being the last thing you almost didn't ship because it felt unfinished?

[Felipe Fumero @FelipeFumero663]

@_x019_ Building your own game engine before building the game on it is the kind of decision most people talk themselves out of. What ended up being the hardest part of the engine itself, rendering, physics, or just the tooling around getting content into it?

[Felipe Fumero @FelipeFumero663]

@BuiltByNav A self-hostable NotebookLM alternative outputting flashcards, quizzes, and mock exams from raw PDFs is more feature-complete than most clones. What's generating them, one LLM pass per format, or a shared extraction step feeding multiple generators?

[Felipe Fumero @FelipeFumero663]

@tualatrix Moving a long-running Docker workload off a Mac mini to a dedicated Ubuntu box to free up memory is a migration people put off. Trickiest part, the service migration itself, or re-pointing everything that depended on the old host?

[Felipe Fumero @FelipeFumero663]

@ItsBarmanji Judge0 self-hosting with the cgroup v2 support gap is a known pain point that trips up a lot of people, glad the community image saved you. What ended up being different in that image's setup, was it patching the isolate sandbox directly, or working around it another way?

[Felipe Fumero @FelipeFumero663]

@PrasiddhNaik A local dev-stack live map with zero SDK, config, or account, and nothing leaving the machine, is the right constraint for a tool people will actually trust to run. How's Runscape discovering Docker containers and ports, polling the Docker socket directly, or something else?

[Felipe Fumero @FelipeFumero663]

@SaiyamPathak An SRE copilot returning 'registry returned 404 for that tag' instead of a vague guess is the diagnostic precision that makes AI-in-ops trustworthy. Is it just kubectl describe plus registry calls chained, or something more structured?

[Felipe Fumero @FelipeFumero663]

@jiahong__chen Building a workaround for Garmin not exposing a public workout-creation API is exactly the kind of reverse-engineering most people give up on. How are you actually authenticating and posting the draft, did you have to reverse Garmin Connect's internal API calls?

[Felipe Fumero @FelipeFumero663]

@vibsbuild A CLI that visually flags prompt injection, including base64 and zero-width smuggling, live and fully local is genuinely useful. What's catching the zero-width characters, input normalization first, or raw byte scanning?

[Felipe Fumero @FelipeFumero663]

@Ashf03 Searching Spotify by vibe via lyric embeddings and FAISS fits how people actually think about music better than tags. How's the matching holding up on instrumental tracks with no lyrics to embed?

[Felipe Fumero @FelipeFumero663]

@nimendra_ @pidotdev Indexing an Obsidian vault into ChromaDB for semantic search instead of reading whole files is the right scoping. How are you handling re-indexing on note changes, file watching, or manual re-run?

[Felipe Fumero @FelipeFumero663]

@kaisercoder A CLI task manager in Go with JSON persistence is a solid first real project for learning a language, small enough to finish, real enough to hit actual edge cases. What tripped you up first, file I/O error handling, or structuring the command parsing itself?

[Felipe Fumero @FelipeFumero663]

@LandseerEnga A vision-based CLI giving Claude Code eyes and hands, no element IDs or scripts, is different from the usual accessibility-tree scraping. What's doing the vision interpretation, a multimodal model on screenshots, or an intermediate layer?

[Felipe Fumero @FelipeFumero663]

@amitwt Breaking a torrent client into bencode parsing, tracker communication, and peer handshakes is the right decomposition. Which part was most fiddly to actually implement?

[Felipe Fumero @FelipeFumero663]

@siddhantangore 8 seconds down to 5 just from making the writes concurrent is a real, measurable win for a first optimization pass. Did you parallelize all 11 writes at once, or batch them into smaller concurrent groups to avoid hammering the API?

[Felipe Fumero @FelipeFumero663]

@itz_sharmaji001 Spotting the missing lock in Broadcast() before it becomes a race condition in prod is the kind of catch that only comes from actually reading the code line by line. Was this your own project's code, or did you find it reviewing someone else's?