AI engineer & researcher

Minds.
Machines.
The math between.

I’m Kipngeno Koech. I build intelligent systems that see, reason, and act — from robotics and foundation models to brain-computer interfaces.

Explore my workLet’s talk
GitHubLinkedInKigali, Rwanda
A little theory. A lot of building.
Portrait of Kipngeno Koech
Currently at Neotix RoboticsMember of Technical StaffVision-language-action models
Robotics & embodied AIAgentic systemsComputational neuroscience

Selected work

Ideas, made tangible.

All projects
Weightless model architecture explorer interface
Developer tools

Weightless

Explore the architecture of massive AI models—without downloading their weights. Built for the browser and VS Code.

  • JavaScript
  • VSCode Extension
  • Safetensors
  • GGUF
  • Hugging Face
Explore project GitHub
More details & ways to try it

A VSCode extension that inspects ML models without loading the weights — header-only reads open multi-GB .safetensors and .gguf checkpoints instantly, even straight from the Hugging Face Hub via HTTP Range requests (a 500 GB model costs a few hundred KB of traffic). Renders branch-and-merge architecture flowcharts with input/output types derived from weight shapes, reconstructed PyTorch-style module definitions, backbone fingerprinting (≈ Gemma 3, ≈ SigLIP So400m), and memory/finetune estimates. Live on the VS Code Marketplace and Open VSX.

Try it in the browserHugging Face SpaceVS Code MarketplaceOpen VSX
Deep learning · from scratch

Deep Learning Libraries (from scratch)

Neural networks, built from the math up. A collection of NumPy libraries with hand-derived forward and backward passes.

  • Python
  • NumPy
  • SciPy
  • Cython
  • PyTorch
  • PyPI
Explore the collection · 4 projects

A suite of four deep learning libraries built entirely from scratch in NumPy (no PyTorch or TensorFlow), each with hand-derived forward and backward passes and published on PyPI as Cython-compiled binary wheels (Python 3.9–3.12; Linux, macOS, Windows).

  • npmlp-core

    Modular MLP framework with linear layers, six activations (ReLU, Sigmoid, Tanh, GELU, Swish, Softmax), Batch Normalization, and vectorized optimizers.

    Read project GitHub Explore

  • custom-cnn

    CNNs from scratch: Conv1d/2d, transposed convolution, max/mean pooling, up/down-sampling, and 7 activation functions.

    Read project GitHub Explore

  • custom-rnn

    RNN & GRU cells with full BPTT, the CTC forward-backward algorithm, CTC loss, and greedy/beam-search decoders.

    Read project GitHub Explore

  • custom-transformer

    Multi-head attention with hand-derived gradients; a Pre-LN encoder-decoder supporting CTC + cross-entropy ASR and decoder-only language modeling.

    Read project GitHub Explore
Agentic systems

Phoenix: Agentic Software Engineering

Multi-agent systems that investigate, reason, and build. Exploring safer, more capable AI-assisted software engineering.

  • Python
  • Multi-Agent LLMs
  • LangChain
  • FastAPI
  • GitHub API
  • SWE-bench
Explore the collection · 3 projects

A family of multi-agent LLM systems for autonomous software engineering, from retrieval-augmented refactoring through safe, end-to-end GitHub issue resolution.

  • Phoenix: Safe GitHub Issue Resolution (IEEE IRAI 2026)

    Accepted at IEEE IRAI 2026. Six specialized agents resolve GitHub issues end-to-end behind a webhook state machine, with seven layered safety controls and baseline-aware test evaluation; oracle-resolves 75% of a SWE-bench Lite slice. Deployed always-on and released on PyPI.

    Read project GitHub Explore

  • Phoenix Agent

    Autonomous code-analysis agent running a 7-phase control loop with human-in-the-loop approval, parallel CoderAgents, WebSocket streaming, and a 3-layer memory system (Redis + PostgreSQL + Neo4j).

    Read project GitHub Explore

  • Phoenix RAG

    Retrieval-Augmented Generation for code refactoring with a ReAct reasoning agent, ChromaDB semantic search, and groundedness verification to reduce hallucinations.

    Read project GitHub Explore

Published Papers

Peer-reviewed research across multi-agent AI, brain-computer interfaces, and power systems: IEEE IRAI 2026, SiPS 2025, and IMAS 2025.

Read more

VLAKit

A modular, config-driven toolkit for fine-tuning Vision-Language-Action models (π0.5, MolmoAct) across ephemeral GPU boxes — crash-resilient auto-resume, FSDP/DDP scaling, and verified W&B weight egress. Read the docs.

Read more

Experience

Where I’ve made an impact.

Research, engineering, and the people I’ve built with.

Full experience

Neotix Robotics

Jun 2026 - Present

Current

  1. Member of Technical Staff - VLA

    Sep 2026 - Present

    Building training, post-training, and evaluation infrastructure for vision-language-action (VLA) and robotics foundation models.

    Role details

    Researching reward modelling, reinforcement learning, recovery, and learning from corrective interventions, with work on robotic perception, embodiment transfer, and real-time policy execution.

    • VLAs / VLMs
    • Robotics
    • Reinforcement Learning
    • LeRobot

  2. Research Intern - VLA

    Jun 2026 - Aug 2026

    Worked in an integrated research and engineering team developing robotics data collection and model-training systems.

    Role details

    Fine-tuned MolmoAct 2 and π0.5 on in-house teleoperation data and ran inference on bimanual YAM robots using LeRobot and Hugging Face tooling.

    • VLAs / VLMs
    • Robotics
    • LeRobot
    • Hugging Face

IEEE Rwanda Sub-Section

Dec 2025 - Present

Current

  1. SAC Chair

    Aug 2026 - Present

    Chairing the Student Activities Committee (SAC) for IEEE Rwanda Sub-Section, supporting student branches and coordinating student engagement and collaboration across universities.

    Role details
    • Leadership
    • Student Activities
    • Community

  2. Treasurer

    Dec 2025 - Present

    Serving as Treasurer of IEEE Rwanda Sub-Section, managing sub-section finances.

    Role details
    • Financial Management
    • Leadership
    • Community

Logic & Matter Labs

May 2023 - Present

Current

  1. Founder

    May 2023 - Present

    Founded an initiative mentoring K-12 students in web development through hands-on workshops at schools.

    Role details

    Founder

    • Education
    • Mentorship
    • Web Dev

Carnegie Mellon University Africa

Sep 2024 - May 2026

  1. Graduate Teaching Assistant

    Sep 2024 - May 2026

    Teaching Intro to Deep Learning, Computational Materials Science, and Bridge Program.

    Role details

    Guiding students through PyTorch, EEG/PPG analysis, and ML-based materials modeling.

    Smart Africa Scholar

    • Deep Learning
    • PyTorch
    • Materials Science

  2. Graduate Research Assistant

    May 2025 - Aug 2025

    Multimodal Biointerfacing & Catalysis research.

    Role details

    Performed computational analyses linking material structures to experimental outcomes using Python and CrystalMaker.

    Summer Research

    • Research
    • Python
    • Materials Science

Research & publications

Questions worth pursuing.

All publications
2026Preprint

Zero-Shot Neural Priors for Generalizable Cross-Subject and Cross-Task EEG Decoding

arXiv preprint · Signal Processing (eess.SP)

Baimam Boukar Jean Jacques, Brandone Fonya, Nchofon Tagha Ghogomu, Pauline Nyaboe, Kipngeno Koech

arXiv:2606.23706DOI
Abstract

A zero-shot cross-subject framework for generalizable EEG decoding on the large-scale Healthy Brain Network dataset, benchmarking a CNN baseline, a hybrid LSTM, and a Transformer-based foundation model. To adapt the Transformer for regression without catastrophic forgetting, we propose a novel progressive unfreezing strategy. The fine-tuned Transformer reaches an nRMSE of 0.9799 on unseen subjects (vs. 0.9991 for the baseline), advancing scalable, calibration-free EEG decoding for computational psychiatry and behavioral prediction.

  • Brain-Computer Interfaces
  • EEG
  • Foundation Models
2026Accepted

Phoenix: Safe GitHub Issue Resolution via Multi-Agent LLMs

IEEE International Conference on Responsible Artificial Intelligence (IRAI), Melbourne, Australia

Kipngeno Koech, Muhammad Adam, Baimam Boukar Jean Jacques, Joao Barros

arXiv:2606.20243Code & Docs
Abstract

A multi-agent LLM system that resolves GitHub issues from triage through pull-request creation, combining seven layered safety controls with a baseline-aware test evaluation strategy. Work is decomposed across six specialized agents (planner, reproducer, coder, tester, failure analyst, and PR agent) coordinated by a label-based GitHub webhook state machine, with every change checked against a baseline test run before a PR is opened. Phoenix oracle-resolves 75% of a SWE-bench Lite slice with no pass-to-pass regressions, and preserves correctness on 100% of a 42-issue pilot across 14 repositories.

  • Multi-Agent LLMs
  • Software Engineering
  • SWE-bench
2025Published

Improving SSVEP BCI Spellers with Data Augmentation and Language Models

IEEE Workshop on Signal Processing Systems (SiPS), Hong Kong

J. Zhang, R. Zhang, K. Koech, D. Hill, and K. Shapovalenko

arXiv:2412.20052IEEE XploreDOI
Abstract

A hybrid framework integrating domain-specific EEG data augmentation with a language model to improve SSVEP-based Brain-Computer Interface speller accuracy for individuals with motor disabilities. It decodes scalp-recorded EEG to identify the characters a user gazes at, tackling high EEG variability and poor generalization to unseen subjects with an end-to-end PyTorch training and evaluation pipeline.

  • Brain-Computer Interfaces
  • EEG
  • NLP

The person behind the work

Curiosity, from first principles.

I’m a researcher and engineer drawn to a simple question: how do we make machines understand and interact with the world?

My work brings together robotics, vision-language-action models, agentic AI, and brain-computer interfaces. Underneath it all, I care most about the math — the linear algebra, probability, and optimization that make the engineering work.

Outside the lab, you’ll find me hiking, traveling, reading, or contributing to the IEEE community.

Away from the keyboard

Education

A foundation for the work.

Full education
Apr 2024 - May 2026

Master of Engineering - Artificial Intelligence

Carnegie Mellon University Africa

Kigali, Rwanda

Details & coursework

Smart Africa Scholar. Graduate TA for Introduction to Deep Learning (11-785) and Bridge Program. Research in computational neuroscience and BCI.

  • Deep Learning
  • Machine Learning
  • Neural Networks
  • Computer Vision
Aug 2025 - Dec 2025

Master of Science - Engineering Artificial Intelligence

Carnegie Mellon University

Pittsburgh, PA

Details & coursework

Advanced study in AI engineering, focusing on building production-ready AI systems and research methodologies.

  • Advanced ML
  • AI Systems
  • Research Methods
Sep 2020 - Dec 2024

Bachelor of Science - Software Engineering

Multimedia University of Kenya

Nairobi, Kenya

Details & coursework

Best Club of the Year 2022. Class Representative throughout. Vice Chair of MMU Tech Community. Foundation in software engineering and algorithms.

  • Software Engineering
  • Algorithms
  • Data Structures
  • Systems Design

Away from the code

Notes on being human.

Essays, reflections, and things I’m thinking about.

ReflectionMar 2026

Paul to the Church in Nairobi

A heartfelt, letter-style reflection addressed to the church in Nairobi — honest, raw, and personal.

Substack · 3 min read
EssayJan 2026

I Want to Disappear

On the urge to vanish from it all and leave only the occasional Substack article as a trace in the world.

Substack · 3 min read
StoicismMar 2025

The Five Good Roman Emperors

After the bloody end of Domitian's reign, Rome entered its golden age. A look at the five emperors who steadied the empire and the Stoic ideals that guided them.

Medium · 4 min read
Substack Medium

Good work starts with a conversation

Have something
in mind?

I’m always up for a conversation about research, robotics, or an interesting problem worth solving.

Get in touchkip@kipngenokoech.com