Ali Kheirkhah

Fullstack Engineer

University of Cambridge architecture
University of Cambridge

What I Built During My Research Internship at the University of Cambridge X Google DeepMind

This summer, I had the opportunity to work as a research intern at the University of Cambridge on a programme powered by Google DeepMind. The focus was not simply theoretical AI research, but building real tooling around language models — infrastructure that makes research measurable, reproducible and visual.

This internship was genuinely one of the best experiences I have ever had. It pushed me technically, strengthened my research foundations, and reshaped how I think about AI systems.

My work centred around Pico-LM, an open initiative exploring small, transparent language models — and more specifically, building the ecosystem around it.

Supervision & Mentorship

I was supervised by Richard Diehl Martinez, Research Fellow in Computer Science at the University of Cambridge and founder of Pico.

Richard played a massive role in shaping my understanding of:

  • Natural Language Processing
  • Large Language Models
  • Research methodology
  • Evaluation and reproducibility

Beyond technical guidance, he helped me develop a deeper intuition for how language models learn, how experiments should be structured, and why transparency matters in research. The mentorship was rigorous but incredibly supportive — and it elevated my thinking far beyond implementation alone.

The Vision Behind Pico-LM

PicoLM.io is built around a simple but powerful principle:

Language model research should be transparent, measurable and accessible.

Rather than treating models as black boxes, Pico-LM emphasises:

  • Reproducibility
  • Clear training metrics
  • Open evaluation pipelines
  • Lightweight experimentation

Instead of scaling blindly, the philosophy is precision over brute force.

Building the Pico-LM Dashboard (Next.js)

My primary responsibility was working on the Pico-LM Dashboard, built using Next.js.

The goal of the dashboard was to make model training and evaluation interpretable at a glance. Research is only useful if you can clearly see what is happening.

What I Worked On

  • Architecting the frontend using Next.js (App Router)
  • Designing clean visualisations for:
    • Training loss curves
    • Evaluation benchmarks
    • Dataset statistics
  • Implementing modular experiment tracking components
  • Optimising performance for large metric payloads
  • Ensuring a clean separation between research data and UI logic

The stack emphasised:

  • Next.js
  • Type-safe APIs
  • Modular component architecture
  • Clean UI patterns inspired by research tooling rather than marketing dashboards

The dashboard became a central interface for:

  • Monitoring model runs
  • Comparing experiments
  • Generating structured reports
  • Debugging training behaviour

In research, clarity beats complexity. That principle guided every design decision.

Contributing to pico-report

Alongside the dashboard, I worked on the open-source library:

👉 https://github.com/pico-lm/pico-report/

Initially, pico-report was created to generate structured experiment data specifically for the Pico-LM dashboard — acting as the reporting layer that fed consistent, reproducible metrics into the visual interface.

However, as the tooling matured, it became clear that the abstraction was useful beyond the internal dashboard. It was then open-sourced so that anyone could use it to structure experiments, generate evaluation outputs, and produce transparent reports.

pico-report is designed to:

  • Structure experiment outputs
  • Standardise evaluation metrics
  • Generate reproducible reports
  • Make benchmarking transparent

Why This Matters

One of the biggest issues in AI research is inconsistent reporting. Results often depend on hidden configurations, dataset variations or undocumented adjustments.

pico-report addresses this by:

  • Defining structured output formats
  • Making experiment metadata explicit
  • Creating reproducible evaluation artefacts

Instead of:

“Trust us, it works.”

The standard becomes:

“Here is the exact experiment, configuration and output.”

Working on this library required thinking not just as a frontend engineer, but as someone designing research infrastructure.

Engineering Challenges

1. Handling Large Metric Streams

Training logs scale quickly. Rendering them naïvely leads to performance issues.

Solutions included:

  • Client-side data chunking
  • Memoised visual components
  • Efficient state management patterns

2. Designing for Researchers

Researchers do not want:

  • Overly stylised dashboards
  • Marketing-heavy design
  • Distracting UI

They want:

  • Raw data clarity
  • Accurate comparisons
  • Fast iteration

Designing for this audience required discipline and restraint.

3. Reproducibility as a First-Class Citizen

Every interface decision had to support:

  • Transparency
  • Auditability
  • Replicability

The UI was not decoration — it was part of the scientific workflow.

What I Learned

AI Infrastructure > AI Hype

The most impactful work often lies not in scaling models, but in building:

  • Evaluation pipelines
  • Data handling systems
  • Reporting frameworks
  • Developer tooling

Without infrastructure, models are noise.

Systems Thinking in Research

This internship reinforced that research engineering is systems engineering.

Training, logging, reporting and visualisation must integrate cleanly. Each layer influences the others.

Mentorship Accelerates Mastery

Working under Richard’s supervision accelerated my understanding of NLP and LLMs in a way that self-study alone could not. Exposure to research-grade thinking changed how I approach experimentation and evaluation.

The Bigger Picture

Being immersed in a Cambridge research environment powered by Google DeepMind meant operating in a culture of:

  • Rigorous thinking
  • High standards
  • Open intellectual debate
  • Deep curiosity

It pushed me beyond product engineering and into research-oriented systems design.

Final Reflection

This internship was not about building flashy AI demos.

It was about building:

  • Tools that make research measurable
  • Interfaces that make models interpretable
  • Libraries that enforce reproducibility

From architecting the Next.js Pico-LM dashboard to contributing to and open-sourcing pico-report, the work focused on making AI research more structured and transparent.

It remains one of the most formative and rewarding technical experiences I have had — and it fundamentally shaped how I think about responsible AI development.

Explore

  • 🌐 Pico-LM Website: https://picolm.io
  • 📦 pico-report: https://github.com/pico-lm/pico-report/

Questions I’m Thinking About

  • How can research tooling become as polished as consumer software?
  • Can reproducibility become the default rather than an afterthought?
  • What does responsible AI infrastructure look like at scale?

The answers likely lie in better systems — not just bigger models.