Entrepreneur / technical builder

I engineer operational friction into systems people can run.

I am an entrepreneur who investigates messy technical problems, defines a practical path through them, and builds the software needed to make that path usable.

See the work Connect

Built around real constraints

Technical prototypes that run

Systems, not slideware

Deployed systems

Case studies in engineering operational flow.

03 / 03

01Technical case studyWorking prototype

Local-first autonomous intelligence

ULTRON

ULTRON is a local-first framework for personal AI agents, built for people who want autonomous intelligence without renting it from someone else's cloud. Agents run on your own hardware with shared primitives for on-device tool use — chat, deep research, scheduled digests and long-horizon monitoring — while a learning loop keeps improving them from local trace data. Every run is evaluated against energy, latency and cost budgets, so the cloud is called only when it is truly necessary. A live dashboard meters the watts and dollars saved by staying local, and agents hold persistent memory across scheduled and continuous runs. The result is personal AI that answers to you — not to an API bill.

01Boot on-device

02Reason with tools

03Meter every watt

04Learn from traces

What it runs: On-demand, scheduled and continuous agents with memory
Core loop: Boot → reason → meter → learn, on your hardware
Designed for: Local-first users who refuse cloud lock-in
In the demo: Chat with tools, deep research and the energy dashboard

Local-firstAutonomous agentsOn-deviceEnergy meteringPersonal AI

Watch live demo

02Technical case studyWorking prototype

Self-hosted intelligence data operations

Morphés

Morphés is a self-hosted, governed training-data operating system for teams that need more than a scraped pile of text. It turns a domain objective into a traceable training handoff: define intent, collect through controlled sources, clean and validate every record, review categories with evidence, then export JSONL only when it is ready to trust. The result is a repeatable, auditable path from research to model-ready data—without surrendering infrastructure or review control.

01Define intent

02Govern sources

03Verify evidence

04Release JSONL

What it delivers: A governed path from intent to model-ready data
Core loop: Intent → collect → validate → review → JSONL
Designed for: Self-hosted teams that need privacy, RBAC and control
In the demo: Guided runs, source evidence, review gates and export

Self-hostedTraceable dataRBACReview gates

Watch live demo

03Technical case studyWorking prototype

LLM fine-tuning visibility

TuneForge

TuneForge is a self-hosted, dark-luxury observability dashboard designed for LLM fine-tuning pipelines. It streams real-time loss curves, GPU hardware telemetry (thermals, VRAM fragmentation), and dataset-integrity signals directly to your own infrastructure. Engineers can act on loss spikes or OOM risks while the run is recoverable, probe fine-tuned models live via an inference playground, and manage model weights in a dedicated checkpoint store — all while tracking estimated GPU costs against a transparent hourly catalog.

01Connect trainer

02Stream signals

03Correlate anomalies

04Ship & probe

What it exposes: Loss, GPU thermals, VRAM drift and dataset insights
When it matters: Correlate hardware health with training performance
How it moves: Live SSE telemetry, checkpoint storage and inference playground
In the demo: Runs, anomalies, hardware health and cost metering

ObservabilityGPU TelemetryCheckpoint storeInferenceSelf-hosted

Watch live demo