20 correctness bugs filed in DataFusion, Arrow, Hyperswitch and Claude Code · fixes merged in DataFusion and Qdrant · 8 PRs in review — since September 2026, each bug with a reproduction, each fix with a test.
Software engineer in São Paulo. Day job: the systems a logistics floor runs on. Open-source time: panics and overflows in Rust codebases, and measuring what LLM agents cost (641 sessions, 134k API calls).
Open to remote roles — Rust, backend, LLM-agent infrastructure. UTC−3: full overlap with US Eastern, 4–5 h with Pacific. English, Portuguese, working Spanish.
Same pattern every time: find the panic, write the reproduction, fix it, keep the test. Combined stars of the projects below: 352k.
| Project | Stars | Merged | Open |
|---|---|---|---|
| apache/datafusion | 9.3k ⭐ | #24916 — free RANGE window frames over Duration/Interval |
13 issues: decimal overflow with negative scale, panics in COPY ... PARTITIONED BY, approx_percentile_cont, Date64 display |
| qdrant/qdrant | 34.7k ⭐ | #10448 — Bits1_5 in the TurboQuant test matrices |
|
| Qiskit/qiskit | 7.8k ⭐ | #16980 — ParameterExpression integer arithmetic overflowed its i64: panic in debug, silently wrong global phase in release. Checked arithmetic in the Rust core |
|
| juspay/hyperswitch | 43.6k ⭐ | #14069 #14070 #14071 — common_utils: masking underflow, MaskedEmail panic on short local parts, Decimal overflow in major-unit conversion |
|
| JuliusBrussee/caveman | 106.6k ⭐ | #1044 — eval isolation, per-language prompts, pt-BR snapshot; measured −8% output tokens in Portuguese | |
| apache/arrow-rs | 3.6k ⭐ | 2 issues — decimal arithmetic overflows i8 precision/scale math for negative scales |
|
| meyfa/CobolCraft | 702 ⭐ | #416 #417 — JSON integer/float encoding bugs, in COBOL | |
| kobotoolbox/kpi | 181 ⭐ | #7586 — save asset and its new version in one transaction (Django) | |
| anthropics/claude-code | 146.3k ⭐ | 2 issues — #94177 prompt-cache forensics from real sessions, with paper-backed mitigations |
📏 LastroWhere the tokens go in a Claude Code session, measured on 641 real sessions / 134k API calls. Cache reads are ~62% of the bill; 80% of cache writes come from a handful of full re-writes (an idle gap past the 1-hour TTL breaks the cache 93–97% of the time). Ships the measurement tools and four hooks that act on the findings. Filed upstream as anthropics/claude-code#94177.
|
Turns a personal library (234 books and papers) into per-chapter markdown with BM25 + local embeddings + cross-encoder rerank and Rocchio feedback, so an agent gets the right 400 words with title, chapter and page. No LLM in the pipeline, no network. Evaluated on 140 questions: the agent choosing the category by hand beats every automatic classifier tried. Docs in EN and PT-BR.
|
|
URL or HTML in, PNG/JPEG/WebP/PDF out, as an API. Paid product with users. GitHub Action and runnable examples in cURL, Node, Python and PHP.
|
A bakery in New York: sell and close the till with no internet, thermal printing, PIN per sale, sync to a web admin (
|
Day job in logistics: several internal systems for the operation on Microsoft Power Apps. The users are the people on the floor, so the constraint is always the same: it has to work on a phone and need zero training.
| System | Problem | Stack |
|---|---|---|
| ERP migration | Delphi 7 + Paradox management system replaced by a current stack | .NET · PostgreSQL |
| Android app on the Play Store | — | Kotlin · Flutter |
| restore-check | Restore a database backup into a throwaway container, validate it, report | Bash · Docker |
