I like taking ideas that feel slightly too ambitious and making them real enough to test, break, and improve. Most of my work sits somewhere between AI systems, applied ML, full-stack engineering, and product design.

• Agentic systems & memory — retrieval, MCP tools, persistent context, multi-provider workflows, and deterministic safety boundaries around probabilistic models.
• Applied ML & data products — satellite/weather fusion, anomaly detection, forecasting, validation, and production APIs around models.
• Full-stack products — real-time state, auth, persistence, WebSockets, offline flows, CI/CD, and the product logic that has to survive outside a demo.
• Interaction-heavy software — voice, accessibility, browser 3D, motion, and interfaces where how the system feels matters as much as what it computes.

• MNEME explores what persistent AI-agent memory should look like when portability, semantic recall, MCP access, attestations, and deletion proofs are treated as infrastructure rather than prompt history.
• Together pushes browser-native 3D much further: a persistent multiplayer life simulator with households, homes, jobs, cooking, NPCs, customization, shared history, and voice foundations.
• SlopeSense combines satellite and weather signals into a probabilistic landslide-risk pipeline with forecasting, retrospective validation, alert generation, and a production API surface.
• Small Merchant applies agents to commerce while keeping payment permission in deterministic backend code, so model behavior cannot bypass user authorization.

• Mike — low-latency Windows voice dictation with Whisper, voice-activity detection, optional LLM cleanup, global hotkeys, and text injection into any app.
• Shooting OS — a focused content operating system for ideation, scripting, planning, research, shooting, offline access, and knowledge-grounded workflows.
• River Watch — Sentinel-1/Sentinel-2 anomaly detection for river monitoring, deliberately designed to surface evidence for human review rather than overclaim conclusions.
• Orator — Spring Boot + React interview-preparation system with structured question generation, timed practice, answer evaluation, auth, persistence, and session history.
• 1ph — hackathon discovery and enrichment pipeline combining Next.js, Python automation, structured extraction, and AI-assisted normalization.
• EvrythingAI — automated AI/tech newsletter pipeline using RSS ingestion, Mistral-based curation/synthesis, GitHub Actions, and serverless delivery.

LLM systems | RAG | MCP | pgvector | agent memory | voice / ASR | geospatial ML | three.js | websockets | CI/CD