This is my current-generation AI work : LLM systems, agentic pipelines, and modular frameworks I'm actively deploying. It's distinct from earlier training-heavy work (StyleGAN, classical CV) because the shift is toward system design: composable AI infrastructure that maintains and improves over time.
Production retrieval on text and image data: hybrid vector search augmented with knowledge graphs. Vectors give semantic match; graphs give structure, provenance, and explainability.
Quality is measured and improved with an automated LLM-as-Judge loop plus human validation : moving beyond brittle keyword search.
Vector DBs · KGs · RAG · evaluation
Outcome: Auditable retrieval that beat keyword-only baselines.
A prompt-based agent that continuously monitors a portfolio and dispatches alerts on violations : across equity, equity derivatives, and currency derivatives.
Most “agents” are chatbots. This is an operational system: it runs unattended and surfaces what needs human attention.
Agentic patterns · Natural Language rules engines · alert infrastructure
Outcome: Monitoring that replaced ongoing manual analyst cycles.
DSPy (Declarative Self-improving Python) : modular AI framework for building agents that can sustain LLM model change and prompr optimization strategies. Given a small amount of real world data any pormpt based agent can be evolved and proted to another LLM with confidance.
Enables reliability and portability across models and inference strategies without rewriting entire pipelines on every model change.
DSPy · Python · pipeline architecture
Status: In progress for active client work.
My AI work has always been at the frontier: 2017–2019 GANs and enterprise vision; 2020–2022 real-time in-browser video AI; 2025+ LLM systems and agents. The tools change; the discipline : shipping AI that works for real clients at scale : does not.
I've shipped production systems. Let's talk about what you're building.
Start a conversation →