✦ Research-Backed Implementation

Reinforcement
Learning that
actually ships

25+ years of RL research applied to real business problems. We bridge the gap between academic theory and production deployment — depth consultancies can't match, velocity academics never deliver.

Typical engagements start at $250/hr · Project work from $50K

🧠
DataWorks Engagement Reinforcement Learning Consulting
Problem scoped & formalized MDP formulation, state/action space, reward signal design
Algorithm selected & benchmarked PPO vs. SAC vs. DDPG — justified for your environment
Model trained, evaluated, deployed With convergence guarantees and failure mode analysis
Team upskilled Your engineers understand what they're maintaining

Hourly rate from $250 / hr
25+
Years of RL Research
Since 1998
70+
Peer-Reviewed Publications
Nature, PNAS, Psych. Review
6,500+
Citations
Google Scholar
100%
Research-Backed Work
No heuristics, no guesswork
Services

What we do for you

Four core offerings, each grounded in formal methodology. Pick one engagement or stack them — we scope everything to fit your timeline and budget.

⚙️

Decision AI Systems

Autonomous decision-making for complex, sequential business processes. Pricing engines, supply chain optimization, adaptive recommendations, resource allocation. Systems that improve over time without manual re-tuning. Our peer-reviewed algorithms mimic decision making in brains.

Sequential Decision-Making Bandit Algorithms Adaptive Optimization
🔬

Data Science Audit

Independent, research-grade evaluation of your existing models and methodology. We find the flaws, quantify the risk, and give you a remediation roadmap. Useful before a major launch, a fundraise, or when something is inexplicably underperforming.

Model Validation Statistical Methodology Pipeline Review
📈

Model Evaluation & Optimization

You have a model. It works, but not well enough. We benchmark it rigorously, identify the failure modes, and optimize performance — whether that means better features, better training, or replacing the algorithm entirely.

Benchmarking Hyperparameter Tuning Architecture Review
The Researcher

Stefano Ghirlanda, PhD

25+ years building the mathematical foundations that other researchers cite — with 70+ peer-reviewed publications cited over 6,500 times.

Computational modeling, neural networks, reinforcement learning, and the formal study of how systems learn from experience. Not as a practitioner following frameworks — as an architect who built some of them.

📄
70+ Peer-Reviewed Publications Nature, PNAS, Psychological Review, Neural Networks
📊
Cited 6,500+ Times Google Scholar verified impact
🧠
25+ Years of RL Research Formal models of associative learning and reward-based optimization
Reinforcement Learning

Formal models of associative learning and reward-based policy optimization. Research that predates and informs modern deep RL — 20+ years of primary contributions.

Computational Modeling

Quantitative models of complex adaptive systems. The same mathematics that describes biological learning also describes how machines should learn from feedback.

Neural Networks

Architecture, training dynamics, and generalization. Research-grade understanding of why neural networks succeed and fail — not just how to run them.

Applied Methodology

Bridging formal statistical theory and real-world data science practice. The difference between models that look good in papers and models that work in production.

Why DataWorks

Research depth.
Implementation velocity.

Most RL consultancies are one of two things: academics who've never shipped production code, or engineers who've never read a paper. DataWorks is neither — and both.

We don't just recommend an algorithm and leave. We stay until it works, explain why it works, and make sure your team can maintain it.

Research Firms
DataWorks
Dev Shops
RL expertise
~
Ships production code
Formal methodology
Understands your domain
~
~
Explains the work
~
Past Work

Selected projects

Research-backed implementations for real commercial problems — performance engineering, model auditing, predictive analytics, and cross-source data engineering.

🔬
Model Auditing

Auditing Inflated Cognitive Claims with a Simpler Model

Bottom Line

Benchmark auditing applicable to LLM evaluation, AI due diligence, and stress-testing "the model can reason" claims.

58 Experiments Audited 88% Variance Explained 14+ Studies
Situation Published literature claimed animals use "intuitive statistics" — a cognitively complex ability requiring high-level reasoning.
Task Independently audit 58 experiments across 14+ studies in animal cognition.
Action Built a minimal associative-learning model to test whether simpler mechanisms could explain the data — without any per-study tuning.
Result Reproduced every result in the literature; explained 88% of variance on average. Claims of "intuitive statistics" were not warranted.
Ghirlanda & Mendoza, Psychological Review (in press) · osf.io/8wjdb
📈
Predictive Analytics

Cultural Trends as Leading Indicators

Bottom Line

Diffuse cultural-choice data as leading indicators for brand sentiment, product reception, and public-figure risk — decades before traditional polling catches up.

135 Years of Records 90%+ Name Collapse 630 Names Analyzed
Situation Traditional polling missed major public-sentiment shifts — evidenced most sharply by the 2016 forecasting failures.
Task Test whether diffuse cultural-choice data encodes public sentiment earlier than polls do.
Action Analyzed 100% of U.S. Social Security name records (1880–2015). Found "Hillary"/"Hilary" collapsed 90%+ after 1992 — the most extreme cycle-skew among 630 comparable names.
Result Effect crossed party lines and was visible in the late 1990s — decades before 2016 polling failures.
Ghirlanda, Cliodynamics 8(1) · doi.org/10.21237/C7clio0033703
🔗
Data Engineering

Cross-Source Data Curation and Unified Modeling

Bottom Line

Fragmented multi-source evidence → defensible schema → single model → actionable verdict. Directly applicable to multi-vendor benchmark audits and cross-study synthesis.

108 Experiments Harmonized 0.88 Mean Correlation 14 Species, 5.9% MAE
Situation Behavioral science evidence is fragmented across dozens of independent labs with incompatible formats and no shared schema.
Task Harmonize 108 experiments and 1,540 data points from 14 species across dozens of labs into a single analyzable dataset.
Action Built per-source ingestion pipelines and one unified 3-parameter model — no per-study tuning, no special-casing.
Result Mean correlation 0.88, mean absolute error 5.9% across 68 datasets.
Ghirlanda, Lind & Enquist (2017), Royal Society Open Science 4: 161011
Pricing

Straightforward rates

No retainer traps. No surprise overruns. We scope tightly and stick to it.

Hourly Consulting

$250/hr
Minimum 4-hour block

Architecture review, technical advisory, model evaluation, or any focused work that doesn't require a full project scope. Good for getting unstuck fast.

Book time →

Project Engagement

$50K+
Fixed-scope, milestone-based

Full RL system design and implementation. We define the scope, agree on milestones, and deliver. Typical engagements run 6–16 weeks depending on complexity.

Get a scope →

Ongoing Advisory

$5K/mo
Monthly retainer, cancel anytime

8 hours/month of dedicated access. Code review, architecture guidance, and a direct line to a 25-year RL expert when your team needs backup.

Start a conversation →
Work With Us

Tell us about your project

We respond within one business day. If the project is a good fit, we'll set up a 30-minute call to scope it properly.

Inquiry received

We'll review your project and respond within one business day.