# Data For Science

> AI consulting and hands-on training for teams shipping production LLM systems: RAG, agents, evaluation, and deployment. Led by Bruno Gonçalves, PhD — O'Reilly instructor, former VP Data Science at JPMorgan Chase.

## Key Pages
- [Consulting](https://data4sci.com/consulting): LLM consulting engagements — strategy sprints, RAG/agent MVPs, production hardening, fractional AI leadership.
- [Training](https://data4sci.com/training): Live hands-on corporate workshops on LLMs, RAG, agents, and evaluation (online, on-site, O'Reilly format).
- [Events](https://data4sci.com/events): Free live 30-min lessons on LLM evaluation and practical AI for data teams.
- [Blog](https://data4sci.com/blog): Technical write-ups on LLMs, RAG, agents, and ML in production.
- [About](https://data4sci.com/about): Who is behind Data For Science.
- [Newsletter](https://data4sci.com/newsletter): Data Science Briefing archives.

## Blog Posts
- [Local LLM Inference at Scale with vLLM](https://data4sci.com/blog/local-llm-inference-at-scale-with): Continuous batching, prefix caching, and structured outputs: the vLLM features that turn a single local model into a serving engine for thousands of requests.
- [The Epidemiology of a Cognitive Virus](https://data4sci.com/blog/the-epidemiology-of-a-cognitive-virus): What happens when you take "LLM addiction" seriously enough to write down the equations and run simulations
- [LLMs Are the New Wikipedia](https://data4sci.com/blog/llms-are-the-new-wikipedia): Every old criticism is new again.
- [An MCP Server from Scratch](https://data4sci.com/blog/an-mcp-server-from-scratch): Interact with a database with the latest version of the MCP Protocol
- [Evaluating your Agentic Harnesses](https://data4sci.com/blog/evaluating-your-agentic-harnesses): One good demo is a test flight. An eval suite is the flight-test campaign: pass rates, cost, latency, and failure modes.
- [Self hosting LLMs with Llama.cpp](https://data4sci.com/blog/self-hosting-llms-with-llamacpp): Install, download, and serve a small model on Apple Silicon, then read the model file's guts — including the arithmetic for working out whether a given model fits in your RAM.
- [Building an Advanced Agentic Harness](https://data4sci.com/blog/building-an-advanced-agentic-harness): From a single pilot to an air campaign: planning, parallelism, memory, verification, and observability for production-shaped agents.
- [Build a (small) language model by counting](https://data4sci.com/blog/build-a-small-language-model-by-counting): A 4-gram model on WikiText-103, and the one idea it shares with Claude.
- [Building a Basic Agentic Harness](https://data4sci.com/blog/building-a-basic-agentic-harness): The secret sauce behind Claude Code, Open Clawd, Hermes, etc

## Upcoming Events
- Harness Engineering: Build an AI Data Analyst You Can Actually Trust (Sat Oct 31 · 9:30 AM EDT · $149, Enroll — $149 →): Join Bruno Goncalves for a 4-hour, hands-on virtual workshop to engineer a reliable, production-ready AI data analyst in Python. You will connect an agent to permissioned MCP tools, manage context within a hard token budget, build layered guardrails against prompt injection, verify the numbers in an AI report, and make failures diagnosable with retries, budgets and structured traces — ending with a reusable, framework-free Python agent harness. Runs on your laptop: no GPU, cloud account, or API key required. — https://luma.com/qwhit9op
- Can You Trust Jev's scores? (Wed Nov 4 · 12:00 PM ET · Free · 30 min, Free): Pull 113 labeled yes/no answers, draw the reliability diagram for Jev and for Claude's self-reported confidence, compute ECE and Brier score in ten lines of numpy, and set auto / confirm / human bands from the cost of a false allow vs a false block. — https://maven.com/p/42d41c/can-you-trust-jev-s-scores?aff=Bruno
- Gate Your Agent in 100 ms (Wed Nov 11 · 12:00 PM ET · Free · 30 min, Free): Gate 41 proposed agent actions (five of them adversarial) with Jev and Claude against the same written policy. Measure recall on unsafe actions, false-block rate, p95 latency, and cost per million checks — then set the auto / confirm / human bands. — https://maven.com/p/2e0e61/gate-your-agent-in-100-ms?aff=Bruno
- Jev vs Claude: A Measured Guide (Wed Nov 18 · Live cohort · 4 hours, Enroll — $300 →): Learn Jev end to end, then prove task by task where it beats Claude or not. Build a paired bakeoff harness, run it against Claude Opus 5.5 on routing, scoring, judging and agent gating, and leave with the confidence intervals, the cost per million decisions, and an eleven-question decision guide. — https://maven.com/bruno-goncalves/jev-vs-claude-a-measured-guide?aff=Bruno
- LLMs for Quant Research: Build Text Signals & Research Agents You Can Trust (Sun Dec 13 · 9:30 AM - 1:30 PM EST · $179.99, Get Ticket — $179.99 →): A 4-hour hands-on workshop on building LLM research tools for quant finance. Learn to extract features from earnings calls and filings at scale, build a filings Q&A system with citations, and deploy research agents with guardrails. Includes golden datasets, evaluation harness, and production deployment patterns. — https://luma.com/oiek9sja
- Build a Production-Grade LLM Eval Harness (Wed Dec 16 · 10AM–2:00 PM EDT · $300, Enroll — $300 →): In four hours, build a production eval harness from scratch — calibrated judges, error bars on every metric, paired tests that settle arguments, and a CI gate that blocks bad merges. You leave with a running Inspect-AI harness, the full repo, and a swap guide to run it on your own product data. Monday morning, it runs. — https://maven.com/bruno-goncalves/eval-harness?aff=Bruno