Flagship theme
Closed-Loop Data Engineering
Systems that detect, evaluate, and improve data pipeline behavior through feedback loops: self-healing pipelines, automated regression detection, and agentic data engineering.
Exploration
LLMs as Data Transformation Engines
How language models become components inside data pipelines. They handle extraction, standardization, and mapping tasks that resist hand-written rules, and their outputs get validated like any other transformation.
Exploration
Knowledge Graphs for Hidden Connections
How entity relationships reveal insights across financial and healthcare data, from AML risk networks to patient identity across fragmented clinical systems. We built a 2.3M-node knowledge graph from the ICIJ Offshore Leaks, OpenSanctions, and news extraction to trace sanctioned money through shell networks.
Exploration
Multi-Agent Development Workflows
Leveraging AI coding with a continuously improving workflow. An agentic team of coding, review, and testing agents works inside the development loop: agents write the first draft, review each other's changes, and run the tests, and an engineer approves what ships. Each iteration feeds back into the workflow, so it gets better with use.
Exploration
Reliable AI Systems
What it takes to run AI in production: evaluation frameworks, monitoring, testing, and workflows that catch model drift and regressions in the pipeline.
Builds Side projects
Working builds where these ideas get tested on real, messy, multi-source data.
Knowledge Graphs · GraphRAG
A knowledge graph to trace sanctioned money through shell networks
A 2.3M-node graph built from the ICIJ Offshore Leaks, OpenSanctions, and news extraction. Walk the traversal from a single UK Treasury fine to the shell-company network behind it.
sources → cleanse → graph db → traversal → context
Marketplace · OCR & Vision
CardTradr: OCR for card trading transactions
OCR and vision models parse photos of graded card slabs, classify each transaction (buy, sell, trade), and match it to grading data. Runs on cloud or local vision models.
Photos → parsing & matching → inventory and P&L dataset
Finance Ops · Billing Pipelines
Billing & invoice automation
Emailed receipts and invoices parsed into amounts, dates, vendors, and GL hints, plus automated monthly invoice generation from time-tracking data.
Email inbox → extraction → validated ledger entries
Want to compare notes?
These are active explorations. Conversations with practitioners make them better.
Start a Conversation