AI / ML engineer building the trust, eval, and runtime layers around LLM agents. I ship small, tested tools that make AI systems measurable and safe to run: mutation-tested evals, guardrails, MCP servers, and local-first agent runtimes.
📍 Philippines (UTC+8) · 🟢 Open to AI/ML engineering roles, remote or on-site
| Project | What it does | Proof |
|---|---|---|
| testteeth | Mutation-test grader for AI-written tests: CLI, pytest plugin, MCP server. Wraps StrykerJS and cargo-mutants for TS/JS and Rust. | On PyPI. Weak AI tests score 20.7% vs 96.6% for strong ones. |
| sealbox | Compute-to-data MCP server: agents query sensitive data inside an encrypted vault, and an egress gate (k-anonymity, DP noise, PII, code fingerprints, leakage budget) checks every output. | 56 red-team attacks, 0 leaks. 344 tests, 95% coverage. |
| semcache | Semantic cache proxy for OpenAI-compatible APIs. | 65.1% hit rate, 0.51% false hits on a paraphrase workload. |
| regresscheck | Prompt/model regression gate for CI: a bad prompt change fails the PR. | Mutation score 95.6%. |
| evalforge | Turns production LLM logs into eval datasets. | Mutation score 81.2% → 98.4%. |
| peertrace | P2P sharing of agent run logs on Hypercore/Hyperswarm: signed append-only logs with live tail. | On npm. |
| project-away | Hands-free PC control: webcam hand gestures drive the cursor, local Whisper handles voice typing. | Windows installer on Releases. |
- AXIOM-AETHER (Rust): test-time-training runtime for fast-weight context adaptation, compression proxying, and MCP access.
- genesis: assurance for machine-authored work. Audit the verifier, not the artifact.
- ghost-chimera-rs (Rust): local agent control plane with provider routing, cost ledger, and an Electron console.
- OpenChimera_v1: local-first multi-LLM orchestration runtime.
- ape-mcp: MCP server with an embedded agent, task graphs, and cost-aware model routing.
- sigil + sigil-index: trust scanning and signed trust badges for MCP servers.
- ChimeraLang + chimeralang-mcp: uncertainty-aware reasoning DSL with probabilistic types, on PyPI.
callinsights (sales call scoring) · tenderbuddy (tender discovery and bid prep) · enterprise-brain (Q&A over company docs) · vitality-bss (healthcare document processing) · clusterx-os (lead qualification agent)
Focus: LLM evals & mutation testing · guardrails & data privacy · MCP servers · model routing & caching · local inference (llama.cpp, Whisper)
- Measure it. Every project ships with numbers you can reproduce, and says plainly what is untested.
- Test the tests. I mutation-test my own suites with testteeth, usually to 90%+.
- Local-first. Runs on your machine or your infra, with clear tool surfaces and no hidden calls.