Data For Science

Ship LLM systems you can measure, trust, and maintain.

The demo was the easy part. The hard part is retrieval that holds up on real queries, agents that don’t wander, and evals that catch regressions before your users do. That is what we build — hands-⁠on, in your stack, with your team — and hand off clean.

No pitch deck. You leave with a written scope and next steps, whether or not we work together.

Schematic of an evals-first engagement A schematic chart, not client data. Thirteen iterations are scored on the same test set and plotted with confidence intervals. The score starts at a prototype baseline, rises through the build phase until its whole interval clears the target that was written down before any code, and from then on the target acts as a regression gate. One later iteration scores below the gate and is blocked before it ships. A mark on the iteration axis shows the handoff: from there your team runs the gate, and the iterations after it still clear it. 04812 Iteration Handoff: your team runs the gate Eval score (arbitrary units) 1 Diagnose2 Build3 Harden Prototype baseline Target, written downbefore any code Regression gate Gate Caught beforeusers see it Schematic of an evals-first engagement A schematic chart, not client data. Thirteen iterations are scored on the same test set and plotted with confidence intervals. The score starts at a prototype baseline, rises through the build phase until its whole interval clears the target that was written down before any code, and from then on the target acts as a regression gate. One later iteration scores below the gate and is blocked before it ships. A mark on the iteration axis shows the handoff: from there your team runs the gate, and the iterations after it still clear it. 024681012 Iteration Handoff: your team runs the gate Eval score (arbitrary units) 1 Diagnose2 Build3 Harden Prototypebaseline Target, written downbefore any code Regression gate Regression caught,before users see it Schematic of an evals-first engagement A schematic chart, not client data. Thirteen iterations are scored on the same test set and plotted with confidence intervals. The score starts at a prototype baseline, rises through the build phase until its whole interval clears the target that was written down before any code, and from then on the target acts as a regression gate. One later iteration scores below the gate and is blocked before it ships. A mark on the iteration axis shows the handoff: from there your team runs the gate, and the iterations after it still clear it. 024681012 Iteration, each scored on the same test set Handoff: your team runs the gate Eval score (arbitrary units) 1 Diagnose2 Build3 Harden Prototypebaseline Target, written down before any code Regression gate Regression caught here,before your users see it
Fig. 1. An evals-first engagement, drawn as a schematic (not client data). Every iteration is scored on the same test set; bars are confidence intervals. Once the target is met it becomes a gate that no later change may fall below.

What we build

  • Retrieved Reranked Top-k cut

    RAG & Retrieval

    Retrieval that holds up on real queries: chunking, hybrid search, reranking, and the eval set that proves it works before it ships.

  • Plan Act Observe Blocked

    Agents & Workflows

    Agents with boundaries: tool design, planning patterns, guardrails, and explicit handling of the ways they fail.

  • Eval score, over time Gate Flagged

    Evaluation & Observability

    Quality, latency, and cost you can actually see: eval harnesses, regression gates, and monitoring your team keeps using after we leave.

Bruno Gonçalves, PhD — founder of Data For Science

Bruno Gonçalves, PhD

Founder & AI Consultant

The common thread: taking LLM, RAG, and agentic systems from prototype to something you can actually run in production.

More about Bruno

  • VP, Data Science & Finance

    JPMorgan Chase

    Led ML and LLM deployments for the bank’s finance divisions before founding Data For Science.

  • 170+ live sessions

    O’Reilly, Pearson & Packt

    50+ courses on RAG, agents, evaluation, and LLMOps, taught to working engineers since 2019.

  • 340+ weekly issues

    Data Science Briefing

    Read by 3,000+ practitioners every week since 2019.

  • PhD, Complex Systems

    Former NYU Data Science Fellow

    Research covered by MIT Technology Review, The Economist, Wired, and The Atlantic.

Engagements that match where you are

Every engagement ends with something your team owns — working code, an evaluation set, and a plan — not a deck.

Engagements: duration, what your team owns at the end, and when each fits best
Engagement Duration, in weeks What your team owns at the end Best when
LLM Strategy Sprint 2 weeks
  • The two or three use-cases worth building first, with ROI framing
  • Data readiness and risk review: privacy, security, compliance
  • An architecture roadmap your engineers can start on Monday
you have budget and ambition, but no agreed first step
RAG / Agent MVP 4–6 weeks
  • A working system on your data — not a slide
  • An evaluation harness and test set you own
  • Deployment plan and clean handoff to your team
the prototype exists and “production” is the open question
Production Hardening 2–4 weeks
  • Eval gates and regression tests, so changes stop breaking things
  • Latency and cost down, with the numbers to prove it
  • Guardrails for the failure modes you’ve seen — and the ones you haven’t yet
it’s live, and every change feels like a gamble
Fractional AI Lead Monthly ongoing
  • Weekly architecture and code reviews
  • Model and vendor decisions made on evidence, not hype
  • Hiring and upskilling, so you need us less over time
you need senior judgment on tap without a full-time hire

We work under NDA, follow your security policies, and can work entirely inside your environment when data can’t leave it.

How we work

  1. Diagnose

    Agree on what “working” means

    Use-cases, data, security constraints, and the metrics we’ll be judged on — written down before any code.

  2. Build

    Ship an MVP with evals from day one

    Every iteration is measured against the same test set, so improvement is a number, not an opinion.

  3. Harden

    Make it yours to run

    Monitoring, regression gates, documentation, and training, so the system keeps working after we step back.

Handoff

Your team owns the code, the evaluation set, and the plan.

We publish the code, not just the advice.

Recent hands-⁠on write-ups from the blog — the same patterns we use in engagements:

AI Engineering Cookbook

Coming 2027 from Pearson. 18 hands-on recipes for production AI systems.

Ready to scope a project?

Thirty minutes. We’ll pin down the problem, the success metric, and the shortest path to a working system — and send you the notes.

Book a call

Or email info@data4sci.com