Boris Brodski · AI · since 2018

AI & machine learning

Models built from zero to production — own datasets, own training, own inference, shipped as an application the customer actually uses. Since 2025 also open-weight LLMs on own hardware, LoRA fine-tuning, and the agent harnesses around them.

Zero to hero, alone. Mask detector: dataset collected and labelled, model designed and trained in TensorFlow, real-time OpenCV inference, Gtk app and web monitoring delivered to the customer.
Custom models, not just fine-tuning. Trading models on market data with proper walk-forward evaluation; CUDA, Numba, Colab.
Local LLMs end to end. Self-hosted inference, LoRA fine-tuning, agent harnesses, skills and MCP tools, multi-agent orchestration.
Mathematical footing. Deep Learning Specialization, reinforcement learning, MIT statistics — and a physics minor.

Get in touch

Vision models from data collection to a customer app

The COVID mask detector is the project I point to first because it is the whole loop in one pair of hands. There was no dataset, so I collected and labelled one. The model was designed and trained from scratch in TensorFlow and had to run in real time on commodity hardware, so inference went through OpenCV. The customer needed to configure and watch it, so it shipped with a Gtk configuration application and web monitoring. From first photo to deployed product in a few months, as a one-person project.

Custom models for hard signals

The trading work was the opposite kind of problem: data in abundance, signal scarce. Feature pipelines in PostgreSQL, TensorFlow training on CUDA and Colab, and the evaluation discipline — walk-forward testing, leakage checks — that decides whether a model is real.

Local LLMs, LoRA and agents

Since 2025 the focus is open-weight language models on own GPU hardware: serving with llama.cpp and vLLM, LoRA fine-tuning for domain tasks, and the agent harnesses on top — writing skills, wiring tools and MCP servers, orchestrating several agents over Matrix. The security of that stack is its own topic; see Security & AI security.

What I do for clients

Feasibility and data strategy: is there enough signal, and what would it take to collect it?
Model design, training and evaluation — with the honest number at the end.
Productisation: inference services, desktop or web front ends, monitoring.
Self-hosted LLM deployments: model selection, serving, fine-tuning, agent integration — on hardware you control.

Relevant work

2025 –Architect, developer

Local AI stack — harnesses, skills and LoRA fine-tuning on own hardware

Self-hosted inference (llama.cpp, vLLM) on own GPU hardware, LoRA fine-tuning of open-weight models for domain tasks, and daily work with agent harnesses — writing skills, wiring tools and MCP servers, and multi-agent orchestration over Matrix.

LoRAllama.cppvLLMHermes AgentOpenClawClaude CodeKubernetesPodman
03/2021 – 10/2022Architect, developerGaming industry2 people

AI-assisted cryptocurrency trading

Custom trading models designed and trained from scratch on market data — feature pipelines in PostgreSQL, TensorFlow training on CUDA and Colab, walk-forward evaluation, and the discipline to tell signal from noise.

PythonTensorFlowCUDAGoogle ColabPostgreSQL
10/2020 – 06/2021Designer, developer (one-person project)Startup
COVID mask detector — vision model from data collection to customer app

The full loop in one pair of hands — collecting and labelling the dataset, designing and training the detector in TensorFlow, real-time inference with OpenCV, and shipping a Gtk configuration app plus web monitoring to the customer.
- Own dataset collected and labelled; model designed and trained from scratch
- Real-time inference on commodity hardware with OpenCV
- Customer-facing Gtk configuration app and web monitoring
PythonTensorFlowOpenCVGtkReactNode.js
2026 –Founder, architectOpen source
Torhaus — a secure host for untrusted AI agents

Infrastructure for running coding and research agents on self-hosted models without handing them the keys — the operational side of local AI.
- Threat model built from the 2026 trojaned-open-weights findings (Cohere/Qwen 3.8, "Model-as-a-Trojan")
- Compartments (Zwinger) with data classification, typed one-way channels (Schleuse), a broker that never lets a credential reach the model
- Egress gateway, MCP/tool proxy and model proxy as the single, policy-checked path out
PythongVisorFirecrackerKubernetesOPAeBPFMCP
2025 –Speaker, researcher

AI security — talk and research on trojaned open-weight models

The security side of running models locally — which harness, proxy and egress settings keep a self-hosted model from becoming an exfiltration channel.

Open-weight LLMsllama.cppvLLMagent harnessesegress analysis

Skills in this direction

Security

★★★★☆ AI security — trojaned models 2025–
★★★★☆ Container hardening: gVisor, Firecracker, seccomp, network policy 2025–