Production systems, live demos, and open source.

Search for Latino-led and Latino-focused films and series, in English or Spanish. Hybrid retrieval (BM25 over D1 FTS5 plus bge-m3 in Vectorize, fused with reciprocal rank fusion) across four Cloudflare Workers. A 97-query golden set runs after every deploy (recall@5 0.818), a versioned 70B groundedness judge gates every generated note (auto-approval 33% → 87% after error analysis), and a remote MCP server lets Claude search the catalog. Connect it in claude.ai with latino-canon-mcp.ai-builders-studio-latinx.workers.dev/mcp. $0/month on the free tier, with a runbook built from real production incidents.

Live Cloudflare Worker implementing the Model Context Protocol (MCP). Exposes three tools (time_now, worker_info, and echo) demonstrating MCP server patterns, Workers AI binding, and edge AI deployment. A reference implementation for agentic tool-use architecture.

Autonomous email-summarization agent using Claude Haiku with function calling for tool-based auto-reply. Demonstrates Claude tool-use patterns, serverless deployment to Vercel, and supply-chain auditability through JFrog Artifactory proxying 900+ packages. Includes 5-point DX friction report with actionable fixes.

The portfolio assistant on this site. RAG over a Cloudflare Vectorize corpus plus 8 agentic tools, wrapped in a production LLM-as-judge evaluation layer: a small model scores every reply for relevance and groundedness — nightly against a fixed regression suite and on every visitor 👍/👎 — with results cached by (query, reply) so continuous evaluation adds near-zero cost. The judge scored everything 5/5 for two weeks until per-question hard-case rubrics gave it something to fail on; it now carries matched guards against both fabrication and over-cautious deflection. Human ratings sit beside the machine verdict as a labelled human-vs-machine dataset. Free-tier Cloudflare Worker on Llama 3.3 70B; hardened through a live neuron-quota incident that was quantitatively root-caused and fixed the same night. 280 tests, red-teamed with zero findings. The methodology is public in the ai-vic-eval repo below. Open the chat widget to try it.

Trained a LoRA adapter (rank 8, 16.8 MB) on AI-Vic's own LLM-judged high-scoring replies, across two base models, and A/B tested every version against the un-adapted base with the same judge that scores production traffic. Every adapter regressed: the judge-filtered training data skewed toward confident assertion, so the model learned to always assert and fabricated on questions the corpus doesn't cover. Shipped the base model plus RAG instead. The adapter and full model card stay on Hugging Face as the documented artifact; the write-up on Medium is the point. Reliability by default, impressiveness by choice.

Source for the Latino Canon live demo above: a pnpm/Turborepo monorepo with four Workers, a shared eval package (recall@k, MRR, nDCG@10, groundedness judge, MCP tool-selection eval), eval-checked deploys, and about 330 tests. The README maps every skill to the code and PR that shows it.

The LLM-as-judge evaluation methodology behind AI-Vic: the judge system prompt with the rationale for every clause, the hard-case rubrics, the training-export filters, a sample nightly scorecard, and the full fine-tuning negative-result write-up. Published so the eval work is verifiable to a hiring manager without exposing the private chatbot repo.

Part 1 of a 3-part learning path that swaps the evaluation stack while keeping the same Recall@k / Precision@k / MRR / groundedness core. Fully offline and beginner-friendly: synthetic datasets, embeddings, vector search, keyword vs TF-IDF retrieval, no API dependencies. The most-starred repo in the portfolio, used directly in Techqueria workshops.

Part 2 of the series: production-grade RAG evaluation with Vertex AI embeddings, BigQuery persistence, and Gemini Vision multimodal extraction. Same metrics core, cloud-native stack.

Part 3: a Forward Deployed Engineer's RAG evaluation playbook on Snowflake Cortex embeddings and native compute, no vendor lock-in. Decision frameworks for converting customer constraints (timeline, budget, compliance) into a retriever recommendation, plus regression tracking.

Enterprise AI operating system orchestrating multi-agent workflows with Claude as the default LLM provider. Features a CI-gated evaluation harness running on every commit, a tool-loop safety circuit breaker preventing runaway agent behavior, and strict output validation requiring cited sources or failing at parse time. Demonstrates production patterns for Claude tool-use and agentic safety.

Black-box pairwise evaluation baseline for comparing chatbot versions using LLM-as-judge. Supports fake and edge judge paths, audit metadata, and CI integration, with a roadmap toward conversation-level multi-turn evaluation. The same LLM-as-judge approach AI-Vic runs live against production traffic.

Research lab for inference-performance benchmarking, light fine-tuning experiments, and agentic system behavior, focused on Codex-style AI coding models. Reproducible experiments for AI coding workflows.

End-to-end analytics and evaluation pipeline for AI-assisted coding: developer telemetry simulation, LLM code-generation quality scoring, and causal inference on productivity metrics.

Tutorials and demos for building real-world applications with OpenAI APIs, covering function calling, RAG, embeddings, and multimodal applications through four progressive labs from API basics to production patterns.

Lakehouse-native evaluation framework measuring regional Spanish LLM performance (El Salvador vs Peru) using Delta tables, Spark, and Databricks. Applies Bronze/Silver/Gold data architecture to LLM benchmarking at scale.

End-to-end ML platform built on Kubernetes and microservices, including containerized model serving, automated retraining pipelines, CI/CD for ML, and production monitoring. Stack: Kubernetes, Docker, FastAPI, MLflow.

Distributed ML pipeline predicting flight delays across 30M+ records using MapReduce, Hadoop, and Apache Spark on Databricks. Applied ensemble methods (GBT, Random Forest) with feature engineering on temporal and weather data.

Applied linear regression and TensorFlow to identify socioeconomic and demographic predictors of hate crime rates across U.S. counties. Surfaced statistically significant correlations to inform policy research.

Multi-database recommendation engine using Neo4j (graph traversal), MongoDB (document store), and Redis (caching) to generate personalized store location suggestions at low latency.

Exploratory data analysis on NFL tracking data using Python, NumPy, and Pandas. Analyzed player movement patterns and derived game-level insights from raw positional data.

Designed and executed a research study on movie revenue predictors using OLS regression, hypothesis testing, and diagnostic analysis to identify drivers of box office performance.

Interactive Tableau dashboard reimagining travel data as a visual guide, layering geographic, seasonal, and sentiment data to surface non-obvious destination insights.

iOS app leveraging real-time safety data and ML-based route scoring to recommend safe running routes. Full mobile + backend stack built as UC Berkeley MIDS capstone project.