Keynote Address

From Yesterday to Tomorrow

Data Science, Agentic AI, and the Human in the Driver's Seat

A 15-Year Retrospective Across Commodity Laptops, Physical Edge UAVs, High-Performance Computing, and Modern Foundation Models.

Recep Erol, PhD in AI • Computational Scientist for AI
Founder & Curator, awesome-datascience

Act I: The Genesis

2010: The Spark & The Underpowered Laptop

  • The Underground Stack: Encountered early NumPy, SciPy, and IPython at a local conference.
  • The "Big Data" Storm: MapReduce, HDFS, Zookeeper, and Hive dominated the headlines.
  • The Physical Constraint: My student computer couldn't even run the Hortonworks Sandbox virtual machine.

The Self-Taught Pivot

Turned to early Coursera and Udacity MOOCs. Completed 10+ rigorous courses back-to-back.

"Hardware and frameworks are transient. The appetite for self-directed curiosity is your only permanent asset."
Act I: The Genesis

Open Source as a Global Commons: awesome-datascience

The Chaos of 2013:

Data science resources, papers, and packages were scattered across fragmented mailing lists, blogs, and IRC channels.

The Solution:

Created an open, curated GitHub repository: awesome-datascience.

The Global Impact

  • Tens of thousands of GitHub stars
  • Hundreds of global contributors
  • Keynote talks, podcasts, and global collaborations

Lesson: Sharing knowledge openly builds career antifragility.

Act I: The Genesis

2012: The "Sexiest Job" & Conway’s Triad

[ Hacking / Code ]

Δ

[ Math & Stats ] ———— [ Domain Context ]

Missing Math or Domain = DANGER ZONE

  • The HBR Boom: Data Scientist declared the sexiest job of the 21st century.
  • The Timeless Truth: Code is only one-third of the equation.
  • The 2026 Warning: In the age of LLMs, anyone can generate code. Missing math or domain knowledge makes you dangerously wrong.
Act I: The Genesis

History Repeats: Waves of Automation Panic

Epoch What Was Automated The Panic What Actually Happened
2014 Raw numerical loops → Scikit-Learn "Manual coding is dead!" Productivity surged; focus moved to feature engineering.
2018 Manual parameter search → Cloud AutoML "Modeling is dead!" Focus shifted to data hygiene and metric-to-ROI translation.
2026 Syntax & boilerplate → Agentic AI "Data Science is dead!" You are promoted to Principal Investigator.
Act II: Across the Stack

Layer 1 & 2: Network Sockets to Relational Databases

  • Starting at the Wire: Network programming projects taught packet latency, concurrency, and socket streams.
  • Databases: SQL query plans, schema normalization, and data mining.
"Data does not originate as a clean CSV. It travels across network boundaries and lands on transactional disks. If you don't understand data provenance, you spend your life debugging preventable upstream errors."
Act II: Across the Stack

Layer 3: Physical Edge Systems — Computer Vision on UAVs

Unmanned Aerial Vehicles (Drones)

  • Real-time image processing on low-power embedded processors
  • Unforgiving memory constraints and strict battery milli-watt budgets

The Lesson

You cannot hide behind an infinite cloud cluster. Algorithmic complexity (O(N) vs O(N²)) and sensor vibration noise demand real engineering rigor.

Act II: Across the Stack

Layer 4: High-Stakes Biomedical — Surgical Simulation

  • Medical Image Analytics: 3D anatomical modeling and surgical telemetry analysis.
  • The High-Stakes Environment: In healthcare, a false positive or an uncalibrated confidence interval impacts human lives.

The Epistemic Standard

Statistical optimization is not enough. You need interpretability, counterfactual testing, and domain validation.

Act II: Across the Stack

Layer 5: Enterprise Impact — A $30B Problem at Walmart

  • Enterprise Scale: Moving from academic prototypes to planet-scale supply chains.
  • The Achievement: Tackled a massive $30 Billion operational challenge, resulting in two patents owned by Walmart.

Commercial Translation

A machine learning model is useless if it cannot be operationalized into enterprise balance-sheet value.

Act II: Across the Stack

The Synthesis: Every Layer Compounds

Networks & DBs

Taught me data integrity and latency bottlenecks.

Edge & Drones

Taught me memory ceilings and real-world sensor noise.

Enterprise & Medical

Taught me ROI translation and life-or-death validation.

"No course or project you complete is wasted. Diverse technical layers compound into unshakeable scientific intuition."
Act III: The Scale Frontier

Entering the World of Supercomputers

  • 6 Years in Academic HPC: Deep dive into large-scale computational optimization.
  • The Reality of Petascale Compute:
    • Ultra-low latency InfiniBand interconnects
    • Distributed parallel filesystems (Lustre/GPFS)
    • Multi-node MPI & SLURM scheduling

Beyond Single-Node Python

When datasets scale to terabytes, naive loops die. You must master distributed memory and parallel paradigms.

Act III: The Scale Frontier

Slashing Processing Time: From Days to Hours

Before

Naive, unprofiled data pipelines taking several days per run, stalling scientific discovery.

After

Restructured memory layouts, vectorization, and parallel decomposition: under 2 hours.

"Brute-force cloud compute is expensive and lazy. Algorithmic optimization is an enduring scientific superpower."
Act III: The Scale Frontier

Heterogeneous Computing: Balancing CPUs & GPUs

  • Modern AI is fundamentally heterogeneous:
    • Host CPUs: Complex branching, serial logic, pipeline orchestration.
    • Device GPUs/Accelerators: Massive parallel tensor contractions.
  • The Real Ceiling: PCIe bus transfer latency and memory hierarchy.

Hardware Fluency

You cannot effectively scale modern foundation models if you treat the underlying silicon as a magic black box.

Act III: The Scale Frontier

Green Computing: The Energy-Efficient Computing Mandate

Training frontier models consumes gigawatt-hours of electrical energy.

The New Metric

FLOPs / Watt

Maximizing model intelligence per unit of energy consumed.

  • Quantization (INT8, FP4) and sparse architectures
  • Pruning, distillation, and efficient attention mechanisms
  • Sustainable AI is one of the highest-impact fields of the next decade
Act III: The Scale Frontier

Workforce Development: Democratizing Distributed Systems

Education Moves Science Forward

Software tools are only as powerful as the community that wields them.

  • Led national workforce development workshops in HPC and distributed computing.
  • Trained students, domain scientists, and interdisciplinary researchers.
  • Democratizing high-performance computing is essential for equitable scientific breakthroughs.
Act IV: The Driver's Seat

The Parable of the Alpine Pass

Person A: Prompt-Only Passenger

  • Knows zero physics; never driven manual.
  • Pushes button: "Car, drive to top fast!"
  • Camera blinded by blizzard; black ice undetected.
  • Crashes violently at the hairpin turn.

Person B: Scientist in the Seat

  • Uses cruise control and computerized ABS.
  • Understands traction physics and tire friction.
  • Spots anomalous sensor readings & overrides autopilot.
  • Navigates the pass and reaches the summit.
Act IV: The Driver's Seat

The Clinical Reality: The 99.7% Model That Kills

The Task: Predict cancer patient remission from 50,000 hospital records.

The AI Agent Result: Generates code in 45s, achieves a staggering 99.7% ROC-AUC.

Top Predictive Feature: DISCH_CODE_902

Hospital Clerk: "Oh, that's the billing code we enter into the accounting system when a patient is formally discharged in full remission!"

Target Leakage. The model didn't cure cancer—it memorized an administrative timestamp.

Act IV: The Driver's Seat

Why "AI-Only" Knowledge is Hollow

  • Prompting is a commodity: Anyone can generate a script.
  • The Verification Gap: You cannot audit what you do not understand.
  • An LLM cannot detect:
    • Target leakage & lookahead bias
    • Confounders & Simpson's Paradox
    • Distributional drift in production
"Prompt knowledge alone is hollow. But Foundational Data Science + Modern AI is an unstoppable superpower."
Act IV: The Driver's Seat

The Reactive Revolution: Marimo.io

Legacy Jupyter Notebooks

Out-of-order execution, hidden global states, unreproducible scratchpads. AI hallucinates on stale memory.

Marimo Reactive DAG

Deterministic, pure Python files, Git-friendly, upstream changes automatically propagate downstream. AI is natively contextual.

"Marimo represents the modern archetype: deterministic, reproducible, and built for AI-assisted empirical science."
Act IV: The Driver's Seat

The Dynamic Duo: Data Science × Agentic AI

HUMAN SCIENTIST IN THE DRIVER'S SEAT
Hypotheses Formulation • Causal Directed Acyclic Graphs • Domain Truth • Ethical Verification
↓ Orchestrates & Audits ↓
AGENTIC AI ENGINE & REACTIVE TOOLS
Rapid Code Synthesis • Marimo Execution • Hyperparameter Search • Automated Data Munging

You are not being replaced. You are being promoted to Principal Investigator.

Act V: The Playbook

The Future Skill Matrix: Where to Invest Your Time

Depreciating Skills

  • Memorizing Pandas/SQL syntax flags
  • Writing routine data-cleaning boilerplate
  • Manual hyperparameter grid loops

Appreciating Skills

  • Linear Algebra: Embeddings, SVD, projections
  • Probability: Causal DAGs, Bayesian updating
  • Systems: Relational schemas, GPU memory limits
  • Domain Expertise: Biology, Logistics, Finance
Act V: The Playbook

The Living Portfolio: Beyond the Titanic Dataset

1. Messy Data

Scrape real city transit APIs, sensor streams, or messy government archives. Show you can handle raw reality.

2. Live Reactive Demos

Deploy interactive Marimo or Streamlit apps. Let employers and professors tweak parameters live.

3. Failure Post-Mortems

Publish writeups explaining where your model failed, how you found data leakage, and how you fixed it.

Employers hire forensic investigators who understand edge cases—not tutorial copy-pasters.

Act V: The Playbook

The Outreach Algorithm: High-Signal Networking

Step 1: Read Preprints: Scan arXiv and bioRxiv for papers that genuinely excite you.

Step 2: Clone & Reproduce: Run their code; if you find a bug or edge case, document it.

Step 3: Send an Informed Value-First Email:

"Dear Dr. [Name], I read your paper on [Topic]. I replicated your experiments using a reactive Marimo notebook and noticed this interesting edge case in convergence... Are you accepting research assistants?"
Act V: The Playbook

The Humble Truth: Continuous Learning is Non-Negotiable

  • I earned my PhD in Artificial Intelligence.
  • I lead research as a Computational Scientist for AI.
  • Yet, if I stop learning for six months, I will become obsolete.
"Technology will never stop evolving. Lifelong learning is not a penalty you pay—it is the greatest privilege of being a scientist."
Take the Wheel

Take the Wheel: Drive Into Tomorrow

Do not fear AI. Master the foundations, embrace modern agentic tools, and stay firmly in the driver’s seat.

Join Our Hands-On Workshop:

Interactive Marimo Notebooks: github.com/erolrecep/from-yesterday-to-tomorrow-ds

Curated Ecosystem: github.com/academic/awesome-datascience

Contact: recep.90@gmail.com | @erolrecep