Santiago Lopez
AI engineer building agents, retrieval systems, and internal tools that move from prototype to production — one already serving ~500 monthly users in production.
- agents with tools, memory, and guardrails
- retrieval and knowledge systems for real teams
- product engineering with strong technical taste
/portfolio
Selected work across applied AI, backend systems, and product delivery. The common thread is simple: useful software, shipped with constraints in mind.
-
Booking businesses lose customers when nobody answers WhatsApp after hours. This assistant schedules, reschedules, and cancels appointments in real time — and has been doing it for ~500 monthly users.
Python, FastAPI, Redis, Google Sheets API, Groq LLM API, asyncio, REST APIs, webhooks
Built for production constraints, not demos: real-time availability, safe rescheduling flows, cancellation handling, and WhatsApp-first interaction for non-technical users.
- Runs in production serving ~500 monthly users over WhatsApp.
- Google Sheets as the availability backend, so the business keeps managing its schedule in a tool it already knows.
- Redis-backed conversation state with strict flow handling, so the LLM never freestyles with a real booking.
- Private repo — client contract prevents public code sharing; architecture and key decisions shown in Schedule Agent repo.
-
Cloud engineers lose hours writing CLI commands and API payloads — and one wrong parameter can break infrastructure. A fine-tuned LLM that turns natural language into exact AWS API calls: structured JSON only, no prose, no invented parameters.
Python, Qwen2.5-7B, QLoRA 4-bit, HuggingFace Transformers, PEFT, TRL, FastAPI, MCP, LocalStack, Docker
A Qwen2.5-7B fine-tuned with QLoRA (4-bit) that outputs only the exact JSON for AWS API execution — validated by a safety layer (schema validation, cost estimation, security policies) and running 100% locally against LocalStack or real AWS.
- Fine-tuned on 10,854 curated synthetic examples with noise samples to prevent hallucinations — trained on a single RTX 5070 12GB.
- 100% tool-selection accuracy, 82.8% field-level precision, 0% hallucinations in production (blocked by the safety layer).
- Fully local: no data sent to OpenAI or third parties; the model generalizes to parameter combinations never seen in training.
- Integrations: CLI, REST API, MCP server (opencode/Claude/Cursor), TUI dashboard, and Docker for GPU-free deploys.
-
Production Agent proved the model with one business. Kore is the next step: the same agent core, generalized into a multi-tenant platform serving WhatsApp, Telegram, and web from a single codebase.
Python 3.11, FastAPI, SQLAlchemy, PostgreSQL, Redis, React, Vite, OpenRouter, Docker
Multi-tenant agent platform for customer-facing support workflows: channel orchestration, conversation memory, and guardrails — evolving a system already validated in production.
- Multi-tenant by design: isolated configuration, channels, and data per client on PostgreSQL + SQLAlchemy.
- Provider-agnostic LLM layer through OpenRouter, so model choice is a config decision, not a rewrite.
- Grows out of Production Agent rather than a blank slate — the architecture answers problems already seen with real users.
-
Field engineers in Vaca Muerta need instant answers from dense drilling manuals — and a hallucinated pressure rating is not an option. A RAG pipeline that answers with citations, validated numbers, and a full audit trail.
FastAPI, SQLAlchemy 2.0 (async), Pydantic, Instructor, Next.js, PostgreSQL 16 + pgvector, Groq, multilingual-e5-large, Docker, pytest
Built to prove that RAG for safety-critical industry is not a chatbot with a vector store: a six-stage typed pipeline — classification, hybrid retrieval, reranking, schema-enforced generation, numeric validation, and safety overlays.
- Numeric validation cross-checks every figure against retrieved chunks — the system abstains rather than inventing a spec.
- 4-level traceability (document → page → section → table/figure), as required by Argentine IAPG regulations.
- Hybrid retrieval: e5-large vectors + PostgreSQL FTS fused with RRF, then cross-encoder reranking — built for Spanish prose mixed with English acronyms (BOP, H2S, PPE).
- Structured generation via Instructor + Pydantic: the LLM cannot return malformed output.
-
Satellite methane monitoring over the Neuquina basin (Neuquén, Río Negro, Mendoza, La Pampa) using Sentinel-5P data — anomaly detection with configurable sensitivity for ESG compliance.
FastAPI, React, Vite, Leaflet, SQLite, Sentinel-5P CH4 (Copernicus)
A full-stack geospatial dashboard for methane leak detection from satellite data. Ingests daily Sentinel-5P CH4 files, computes anomaly thresholds (1σ/2σ/3σ), and renders an interactive map with alert zones and 7-day trend charts.
- Real satellite data pipeline: automated Sentinel-5P download, NetCDF parsing, and daily methane column storage in SQLite.
- Configurable sensitivity via sigma slider — users tune false-positive vs false-negative tradeoff for their compliance threshold.
- Interactive map with vector/satellite basemaps, color-coded alert zones (yellow/orange/red), and CSV/PDF report export.
- Deployed on Vercel (frontend) + Render (backend API) — real production deployment for a technical demo.
/blog
Notes on AI systems, engineering judgment, product tradeoffs, and what breaks once real users show up.
Notes coming soon.
/contact
If you are building something ambitious in AI and need an engineer who can move between model behavior, systems design, and shipped product, reach out.
- email — contacto.ssnt@gmail.com
- github — github.com/CodeSantiago
- linkedin — linkedin.com/in/codesantiago