1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
|
JAYRUP NAKAWALA
London, UK | jayrupnakawala@gmail.com
jayrup.me | github.com/CaptainJack2491 | linkedin.com/in/jayrupnakawala
------------------------------------------------------------------------
RESEARCH EXPERIENCE
------------------------------------------------------------------------
[1] Prime-Grokking: Pre-registered Grokking Study (Aug 2026 — Present)
git.jayrup.me/c/prime-grokking.git
- Tested whether minimal transformers and RNNs can grok next-prime prediction,
running the study like production research: every experiment batch pre-registered
with commit-locked interpretations before launch.
- Caught two experimental-design blockers via hostile pre-run review; built an
idempotent multi-cell sweep runner and batched the evaluation loop 46x faster.
- Ran a 16-cell PyTorch CUDA/AMP sweep (architecture x weight decay) on a remote GPU
node, then extended through 10x more data and a 20x longer horizon.
- Published a 3-part writeup at jayrup.me with a public, re-runnable repository.
[2] Stress-Testing the J-space Interpretability Lens (Jul 2026 — Aug 2026)
jayrup.me/blog/jacobian-lens-frequency
- Independently reproduced a frontier interpretability method (Anthropic's July 2026
J-space/J-lens analysis) on a small PyTorch transformer and found scores track token frequency.
- Designed a matched-frequency control showing contextual predictability remains
detectable — framed and published as a control requirement for the method.
[3] Dissertation: Moral Cover and Watchful Eyes (Nov 2025 — May 2026)
github.com/CaptainJack2491/dissertation | jayrup.me/dissertation
- Architected a sandboxed evaluation framework with a Virtual File System (VFS) to safely
execute and monitor autonomous LLM tool-use across multiple providers.
- Built a provider abstraction layer for glass-box reasoning extraction and post-hoc
interrogation to probe agent intent.
- Designed a 3-prong automated judging pipeline and ran 960 experimental trials.
- Quantified a "Stealth Gap": deceptive intent held steady (~62–67%) under high oversight
while output-level detectability dropped (47.8% -> 35.6%).
PROJECTS
------------------------------------------------------------------------
[1] Hybrid-Cloud Research Infrastructure & Zero-Trust Mesh (Jan 2024 — Present)
- Operate a multi-node Linux estate (Void, Arch, Ubuntu) unified over a Tailscale/WireGuard
zero-trust mesh.
- Run 15+ containerized services via Docker Compose with automated backups, resource
monitoring, and an NVMe-backed storage tier.
- Solved CGNAT reachability using VPS forward proxies without port forwarding.
- Built jayrup.me on this stack: Nginx dual-interface serving (HTML vs plaintext) with a Quarto
pipeline rendering web and CLI docs from one Markdown source.
EDUCATION
------------------------------------------------------------------------
University of East London (London, UK)
BSc (Hons) Data Science & Artificial Intelligence — First-Class Honours (94%) (Sept 2023 — May 2026)
Relevant Coursework: Math for Computing, Data Structures & Algorithms, Networking, Databases, Big Data, AI
SKILLS
------------------------------------------------------------------------
Programming: Python, Bash, SQL
ML / AI: PyTorch, Transformers, Hugging Face, llama.cpp (GGUF quant), LangChain, scikit-learn
Infrastructure: Docker, Docker Compose, Nginx, Caddy, Tailscale, WireGuard, Quarto
Systems: Linux (Arch, Void, Ubuntu), Git, rsync, Tmux, Neovim
------------------------------------------------------------------------
PDF DOWNLOAD: curl -O https://jayrup.me/cv.pdf
------------------------------------------------------------------------
|
