Geofrey

NJOROGE

KAMAU.

Computer Scientist from Nairobi, Kenya. I design and build systems from the ground up — research platforms, AI agents with long-term memory, full-stack web applications, and distributed infrastructure. Clean code, real deployments.

View ProjectsRead BlogGitHub

10+

Projects Built

4

Research Posts

2025

CS Graduate


Work

My Portfolio

Research tools and AI agents — built to actually run in the real world.

Kroniqo

● Active

AI agent that "ages" via a SQLite Consequence Graph with recency decay (e^−0.03×days). Multi-backend LLM routing across 7 providers, proactive outreach, sub-agent system, and a FastAPI analytics dashboard with Chart.js.

PythonFastAPISQLiteGroqTelegram

CircuitLens

● Research

Full-stack mechanistic interpretability platform on TransformerLens + FastAPI + Next.js. Key findings: L11·H31 confirmed as Llama-3.2-1B's IOI name-mover head; parametric memory collapses from 98% → 1.6% under adversarial repetition.

TransformerLensFastAPINext.jsPyTorch

Gattai

● Active

Unified AI orchestration layer — combining specialised agents for seamless multi-task execution across complex, multi-step workflows. Designed for scenarios where single agents fall short.

PythonMulti-AgentOrchestration

Dev Blog

● Live

Research blog documenting mechanistic interpretability findings — 4 published posts. Built with Next.js 15, MDX, KaTeX for math rendering, and rehype-pretty-code for syntax.

Next.js 15MDXTailwindCSSKaTeX

Capabilities

Languages & Tools

Frameworks & Libraries

FastAPI90%
Next.js / React85%
TransformerLens82%
PyTorch78%
TailwindCSS88%
SQLite / PostgreSQL80%

Languages

Python95%
JavaScript / TypeScript82%
MQL578%
SQL84%
Bash / Shell85%
C++81%

Tools & Platforms

GitGitHubGroq APIOllamallama.cppCloudflareVercelGoogle ColabTailscaleTelegram Bot APIDiscord WebhooksSupabaseDockerJupyter

Writing

Research Posts

Mechanistic interpretability findings — circuits, memory, and what transformers actually know.

All posts
Mechanistic InterpLlamaIOI

IOI Name-Mover Head in Llama-3.2-1B: L11·H31

Confirming L11·H31 as the IOI name-mover head through 8 experiments including cross-lingual tests with Kikuyu and Swahili names.

Read post

MemoryAdversarialAttention

Parametric Memory Collapse Under Adversarial Repetition

LLM confidence drops from 98% to 1.6% under adversarial repetition — revealing the boundary between parametric and in-context knowledge.

Read post

OPTFactual RecallKnowledge

OPT-1.3b Stores France→Paris in a Single Head (L21·H0)

A single dominant attention head encodes the France→Paris factual association — concentrated versus distributed knowledge across model families.

Read post

SafetyJailbreakCircuits

Jailbreak Token-Copying in Transformer Models

Documenting how jailbreak attempts create token-copying circuits that bypass normal safety mechanisms in auto-regressive transformer models.

Read post

Get in touch

Let's Connect

Open to engineering roles, research collaborations, and interesting problems worth building solutions for — and always collaborating.

Nairobi, Kenya — EAT (UTC+3)
GitHubHuggingFaceEmailDev Blog