Chilean engineer. I work on the layer most people skip: the engine, not the model.
Frontier AI is not limited by intelligence anymore — it is limited by who controls the infrastructure to run it. So I contribute to the projects that shrink that gap: inference engines that fit a dense trunk into consumer RAM, terminal tools that survive real workloads, graph databases that make knowledge queryable without a cloud bill.
I don't open PRs to test the water. 90% of the PRs I open get merged. Every claim below is a commit someone else reviewed and accepted.
IntegrAI — an early-childhood assessment system built in 48 hours, awarded 2nd place at the AWS Breaking Barriers Chile 2026 hackathon — AWS's global hackathon, in its first edition in Latin America.
It wasn't a prototype that died with the demo. It went to production:
- 31 pull requests merged — the most of any project I've touched
- A deterministic + ML decision engine with a margin-based gate and a human-in-the-loop verification queue
- Real-time dashboard (SSE) so teachers see each case as it arrives
- WhatsApp ingestion, multi-child message routing, and a held-out evaluation improved from 11.9% → 96.3%
A system that decides something about a child's development must be auditable, not confident. That principle shaped the architecture.
| Project | What it is | Stars | Merged | What I shipped |
|---|---|---|---|---|
| Colibri | Frontier MoE inference in pure C, zero dependencies | 48.3k | 3 / 3 | TRUNK_RESIDENT_LAYERS=N — dense trunk from 1,128 MB → 64 MB (−94%), byte-identical output. Shipped in the official release. |
| TUIOS | Terminal UI OS / multiplexer in Go | 6.4k | 12 / 15 | Data-race fix that crashed the daemon (#184), clipboard fallback for SSH/VTE, percent-based pane resize, zen mode, regression tests. |
| MillenniumDB | Property-graph + RDF engine in C++ | 267 | 2 / 5 | One-shot installer and server bootstrap scripts (#111, #112). |
| hermes-agent | Agent framework by Nous Research | 261k | in review | samskara — a relational memory provider plugin with weight and doubt. |
66 merged · 48 in repositories I don't own · 73 opened · 90% merge rate.
Measured directly against the GitHub API (Nov 2026). GitHub's Pull Shark tiers lag by months — they count what was merged a while ago, not what shipped last week.
expediente · Python · verifiable evidence engine for public procurement
Every claim is footnoted to a source document. 264 tests, real CI against a MillenniumDB instance, 64.8M triples loaded, 12 topics, MIT. Anti-corruption infrastructure for Chile — built so a journalist can check it, not trust it.
onboarding-hub · TypeScript · a video production studio operated by one person
Screen recording, auto-zoom, FFmpeg render pipeline, Loom alternative, self-hosted. 18 PRs, Apache-2.0, in production. Hardening passes P0→P3: versioned schema, streaming uploads, rate limiting, reproducible deploy from a fresh clone.
hermes-samskara · Python · relational memory for agents — multi-hop recall that knows how much to trust what it remembers.
confucius-agent · Python · three-tier memory for AI agents: mental models → episodes → raw context.
Contributions (12 mo) ..... 374 Merged PRs ............ 66
Active days ........... 50 / 370 Merge rate ......... 90.4%
Longest streak ............ 18 days Repos ................. 10
Best single day ........... 51 Reach (top repo) .. 48.3k ⭐
Also: LLM inference internals · property graphs (SPARQL, Cypher) · agentic systems & memory · CI/CD · computer vision (MediaPipe, YOLO) · satellite ML (ESA / ABOSI4EO)
Open to collaboration on AI infrastructure, agent memory, and public-interest software.
Contributing to open source isn't charity — it's the only way to own the road you drive on.