This repository is a public log of my transition from backend engineering into AI safety and AI engineering.
I spend my days building reliable software — Java, Spring Boot, well-shaped APIs — and my evenings studying what it would take to build reliable AI. I document the work in the open: notes, reading, small experiments, and the things I get wrong along the way.
The intended output is not a portfolio. It is a long-running notebook, kept for years, that records how an engineer actually moves into this field — one paper, one eval, one failed reproduction at a time.
A working history. Intentionally incomplete — milestones get added as they happen.
2024 ──┐
│ Started Software Engineering
│
2025 ──┤
│ Professional Java Backend Engineer
│
2026 ──┤
│ Selected for BlueDot Technical AI Safety
│
Now ──┤
│ Building Kreta
│
Now ──┤
│ Learning AI Safety · Alignment · LLM Reliability
│
Next ──┤
│ AI Alignment Engineering
│
... └─ future entries
A live laboratory notebook. Each topic is a working file under /research. Most entries are preliminary — they grow as understanding does.
Status legend: Reading active study · Drafting summary stable · Working running experiments · Revisited returned to after a gap
| Status | Topic | Summary | Notes |
|---|---|---|---|
| Reading | Prompt Injection | Instruction/data boundary inside LLM systems | placeholders |
| Reading | Mechanistic Interpretability | Reverse-engineering internal model computation | placeholders |
| Reading | AI Evaluations | Capability and safety evals as infrastructure | placeholders |
| Reading | Transformer Circuits | Attention and MLP circuits, induction heads | placeholders |
| Reading | Agent Safety | Failure modes of autonomous LLM loops | placeholders |
| Reading | Scalable Oversight | Supervising stronger actors from weaker ones | placeholders |
| Reading | Constitutional AI | Self-critique against written principles | placeholders |
Topics are added as I start them. Files I have not yet opened are not listed here.
A queue, tracked honestly. Bars show where I am; links go to notes when they exist.
|
AI Safety Fundamentals — BlueDot curriculum
Mathematical Framework for Transformer Circuits — Elhage et al.
In-context Learning and Induction Heads — Olsson et al.
Language Models Can Explain Neurons in Language Models — OpenAI
Constitutional AI — Bai et al., Anthropic
|
backlog — papers queued for later (click to expand)
- Toy Models of Superposition — Elhage et al.
- Sleeper Agents — Hubinger et al.
- Scaling Monosemanticity — Templeton et al.
- AI Safety via Debate — Irving et al.
- Are Emergent Abilities a Mirage? — Schaeffer et al.
- Not what you've signed up for — Greshake et al.
- placeholder — add next paper
Small implementations to test understanding. Most fail or produce noise. That is the point.
| Status | Name | What it tests | Repo |
|---|---|---|---|
| Planning | llm-eval-001 |
A minimal capability + safety eval harness from scratch | coming |
| Planning | induction-heads |
Reproduce induction-head detection on GPT-2 small | coming |
| Planning | inject-lab |
A controlled indirect-prompt-injection testbed against a tool-using agent | coming |
| Planning | weak-to-strong |
Minimal weak-to-strong generalisation experiment | coming |
| Planning | agent-harness |
An LLM agent with an explicit, auditable permission surface | coming |
Empty repos are not created until I have time to start them. Names are reserved.
Short, dated entries. Easy to maintain by hand. Newest on top.
2026-07-17 — Started this notebook publicly. Kreta continues in parallel; research log begins here.
2026-07-15 — Accepted into BlueDot Technical AI Safety.
2026-07- — placeholder
Things I cannot answer yet and am actively thinking about. Listed to show what I am working toward, not what I claim.
- How should we evaluate reasoning rather than retrieval?
- What practical interpretability techniques transfer from small to frontier models?
- How can agent failures — long-horizon, sparse, compounding — be measured before deployment?
- Is instruction/data separation a property we can train, or only scaffold?
- What is the smallest meaningful eval set that predicts deployment risk?
- Can oversight keep up when the actor outpaces the overseer?
- How do we detect when a model has internalised a principle versus learning to recite it?
Added as I notice them. Never removed; downgraded to a Research Log entry when answered.
|
Engineering Java · Spring Boot · System Design · Enterprise APIs · Backend services with clear architecture · Maintainable, observable systems · Developer tooling and DX |
Research AI Safety · Alignment · LLM Reliability · Reliable AI systems · Model evaluations · Mechanistic interpretability |
|
Building Kreta — home for modern software and AI products Early-stage work where backend reliability meets applied AI. |
Learning BlueDot — Technical AI Safety Structured study of fundamentals, alignment, and frontier risks. |
