I find the operational bottleneck that's actually worth solving, and then build the system that removes it.
Ten years turning enterprise problems into working software — most recently at Hilti, where I ship production AI to people who have to use it every day. My open-source work is where I go deeper: retrieval engines, agent workflows, and the guardrails that make either one trustworthy.
MSc in Data Science at Pontificia Universidad Católica de Chile.
📍 Panama City, Panama · 🌐 jasonssdev.com
🧠 OpenKOS · Python Apache-2.0 PyPI
Local-first knowledge engine implementing Google Cloud's Open Knowledge Format. Hybrid retrieval fuses lexical (SQLite FTS5), vector and graph signals through reciprocal rank fusion with PageRank, over a typed knowledge graph built by entity extraction. A fail-closed confidentiality filter guarantees sensitive content never reaches the model, and a reversible forget / irreversible purge lifecycle implements the right to be forgotten down to git history. Runs 100% offline on local models. 19-verb CLI, ADRs, full project governance.
Event-driven expense-approval workflow modelled as a node graph: Pub/Sub ingestion, conditional routing by amount, an LLM reviewer returning Pydantic-structured risk analysis, and real human-in-the-loop interrupt/resume. A custom security checkpoint redacts PII and detects prompt injection before any content reaches the model or the trace log — asserted by integration tests, not assumed. Scored by two purpose-built LLM-as-judge metrics: routing correctness and security containment.
🎙️ Whisper Loopback · Python whisper.cpp Metal
Real-time, fully on-device transcription of macOS system audio through a virtual loopback device. Streaming output via LocalAgreement-2, automatic English/Spanish detection, speaker-labelled meeting mode, optional local summarisation through Ollama. No audio ever leaves the machine.
📰 AI Observatory · Python
Local-first pipeline for daily AI news ingestion and weekly ranked topic synthesis, deliberately scoped to a minimal ingest → rank → synthesize core.
| What | Scale | Outcome |
|---|---|---|
| Conversational AI assistant for Material Management | 300+ users, 5+ countries | ~1,000 specialist hours recovered per year |
| Python + SQL ETL pipeline feeding automated reporting | 300+ sales territories | 15,000+ annual cross-team interactions cut by >50% · CHF 105K/year |
| Regional eCommerce & analytics ownership | 6 LATAM countries | +35% eCommerce sales · +80% new user registrations |
AI & Agents
Data & Platform
Also working with: LangChain · Pydantic · pandas · SQLite/FTS5 · Vertex AI Agent Engine · Cloud Run · Microsoft Copilot Studio · Power Automate · pytest · uv
I help run and build material for PyData Panama — I designed the curriculum, prerequisites and contribution guidelines for the community's Python data-analysis course.
