Artificial intelligence

The most interesting material I've ever built with.

I've been shipping AI products since large models became real — not demos, but systems in production: copilots that whisper in live meetings, agents that finish real workflows, retrieval that cites its sources, voices that hold a phone conversation and video models under genuine creative control.

06 AI builds
Agents · RAG · Voice · Diffusion · Copilots
Local and cloud models
Evals, guardrails and human-in-the-loop by default

AI work

Shipped intelligence, not slideware.

Every build here runs — agents with enforced policies, retrieval with citations, voices with latency budgets, video with creative control.

Private beta

LiveCue AI

AI · Desktop · Real-Time Assistance

An AI meeting assistant that provides contextual live guidance while respecting presentation and screen-sharing workflows.

Challenge: Give speakers useful AI guidance during live meetings — transcribing, understanding context and suggesting responses — without ever leaking the assistant into a shared screen or recording.
Contribution: Built the desktop application: live transcription pipeline, contextual AI response generation across local and cloud models, meeting-platform integration and the privacy-aware overlay system.

Desktop app
Live transcription
LLM integration
Local models
Overlay rendering
Meeting platforms

View case study

LiveCue AI end-to-end system flow: audio input, transcription pipeline, context engine, AI generation layer and a private user-only overlay, closed by a continuous feedback loop.

Concept · in development

Phandor

AI Automation · Productivity

An intelligent desktop automation concept designed to help users complete repetitive digital workflows.

Challenge: Let an AI agent operate a user's desktop to finish repetitive workflows — while keeping the human decisively in control of anything consequential.
Contribution: Designed the agent architecture: context understanding, workflow execution engine, human-approval checkpoints and the safety boundaries that decide what the agent may never do alone.

AI agents
Desktop automation
Workflow engines
Context models
Approval systems

View case study

Phandor end-to-end workflow flow: a user request moves through the perception layer, context model and planner to a human approval checkpoint before the execution engine runs the approved step, with every stage writing to an audit trail.

Private build

AI Video Studio

AI · Generative Video · Platform Build

A Higgsfield-style generative video platform turning prompts and stills into cinematic, motion-controlled clips.

Challenge: Serve GPU-heavy diffusion video generation to consumers at tolerable latency and unit cost — while giving them real creative control over camera motion instead of a prompt lottery.
Contribution: Built the platform end to end: Next.js creation studio, Python inference services, GPU worker orchestration with queueing and progress streaming, motion-preset system, moderation pipeline and credits billing.

Next.js
TypeScript
Python
FastAPI
Diffusion models
GPU workers

View case study

AI Video Studio end-to-end creation flow: user input and motion preset selection lead to a low-cost preview job, a moderation gate before GPU spend, the GPU render pipeline, output review and final delivery over CDN.

Delivered

Knowledge Copilot

AI · RAG · Enterprise Knowledge

A retrieval-augmented assistant that answers from an organization's own documents — with citations, not confidence tricks.

Challenge: Company knowledge lives in thousands of documents nobody re-reads. The copilot had to answer from that corpus accurately, cite its sources, respect who is allowed to see what — and refuse when it doesn't actually know.
Contribution: Built the full RAG system: ingestion and chunking pipeline, hybrid retrieval with reranking, permission-aware search, grounded generation with citations, and the evaluation harness that keeps quality honest.

TypeScript
Python
LLM APIs
Embeddings
pgvector
Hybrid search

View case study

Knowledge Copilot grounded answer flow: document sources are parsed, chunked, embedded and indexed, then a user question runs through hybrid retrieval with ACL filters applied before grounded generation returns a cited answer or says it does not know.

Private beta

Autonomous Support Agent

AI · Agents · Customer Automation

A tool-using support agent that resolves real tickets end to end — and knows exactly when to hand off to a human.

Challenge: Support automation usually means deflection. The goal here was resolution: an agent that can look up orders, apply policies and execute real actions — with mechanical limits on what it may do without a human.
Contribution: Designed and built the agent system: the reasoning loop, typed tool layer, policy engine gating consequential actions, escalation logic, and the simulation-based evaluation rig that replayed historical tickets before anything touched production.

TypeScript
Node.js
LLM tool use
Function calling
RAG
PostgreSQL

View case study

Autonomous Support Agent end-to-end flow: a customer message is understood and planned, context gathered, an action decided against policy, permitted tools executed and the result verified — with a human approval checkpoint for consequential actions and escalation when the agent is uncertain.

Prototype · in development

Realtime Voice Agent

AI · Voice · Real-Time

A low-latency voice agent that holds natural phone conversations — booking, qualifying and answering without the robotic pauses.

Challenge: Voice AI lives or dies in the silences. Every stage — hearing, thinking, speaking — adds latency, and anything past a beat feels broken. The agent had to respond inside conversational rhythm and cope with humans interrupting it mid-sentence.
Contribution: Built the streaming voice pipeline end to end: telephony integration, streaming speech recognition, sentence-pipelined LLM and speech synthesis, barge-in handling, mid-call tool use and the latency instrumentation that made every stage accountable.

Python
Telephony / WebRTC
Streaming STT
LLM streaming
TTS
VAD

View case study

Realtime Voice Agent end-to-end conversational flow: an incoming call streams through speech recognition, LLM reasoning, speech synthesis and playback, with mid-call tools and an always-on turn-taking and barge-in controller that can interrupt any stage.

How I build with AI

Enthusiasm with an engineering spine.

Evals over vibes

Every AI feature ships with an evaluation suite. If quality isn't measured, it isn't real — demos lie, distributions don't.

Guardrails in the engine

Safety, permissions and spending limits are enforced mechanically in the execution layer — never delegated to a system prompt's good intentions.

Latency is the interface

Streaming, partial results and interruption handling decide whether AI feels like intelligence or like a loading spinner.

Humans stay in command

High-stakes actions get approval gates and audit trails. The best agents know exactly what they're not allowed to do alone.

Working surface

LLM product integration
RAG & vector search
AI agents & tool use
Diffusion & generative video pipelines
GPU inference orchestration
Local & cloud model routing
Realtime transcription & voice
Fine-tuning & LoRA workflows
Prompt design & evaluation
AI safety boundaries & moderation

Building something with AI?

From a first LLM feature to a full agentic product — I can take it from idea to production, with the guardrails and evals that make it dependable.

Start a conversation All work