Trust & Safety · Data · AI Automation
I build careful, trustworthy software: systems that keep people safe, turn messy data into decisions, and let machines handle the tedious work. Forever a student of the craft.
Every project here began as something I didn't yet know how to do. I'm a developer who's happiest at the edge of my own understanding, turning curiosity into software that earns trust, makes data legible, and automates the tedious parts away. Three questions pull me forward: Is it safe? Is it true? Can a machine do it for us?
Moderation, fraud detection, and abuse prevention: building the guardrails that let good communities scale.
Pipelines, dashboards, and geospatial insight that turn raw events into the handful of numbers teams actually act on.
Agentic workflows and document intelligence that put LLMs to work on the repetitive tasks, with humans in the loop.
A rotating set of projects across trust & safety, analytics, and automation. Repos and case studies land as each ships.
A multi-modal moderation pipeline that scores text, images, and links against configurable policy taxonomies. Fast heuristic filters catch the obvious cases; an LLM triage layer handles the grey area and escalates borderline content to a human-review queue. Ships with audit logs, appeals, and per-policy dashboards so decisions stay explainable.
Graph-based detection of coordinated inauthentic behavior: bot rings, fake-account clusters, and review fraud. It builds an account-interaction graph, runs community detection to surface suspicious clusters, then ranks them by risk so investigators spend their time where it matters most.
End-to-end event analytics: an ingestion pipeline streams product events into a columnar warehouse, and a dashboard turns them into funnels, retention cohorts, and anomaly alerts. Built to answer the question every team keeps re-asking, what changed, and why, without a data-engineering detour each time.
A location-analytics toolkit that transforms raw GPS traces into meaningful places, trips, and dwell patterns: road-matching noisy tracks, clustering stays into named visits, and visualizing movement over time. Grew out of the mapping engine behind byaristo and generalized into a reusable pipeline.
A framework for chaining LLM agents into reliable, repeatable workflows like ticket triage, report generation, and data enrichment. It handles tool-calling, retries, guardrails, and human-approval gates, so automations stay trustworthy in production instead of quietly going off the rails.
Automates the boring middle of document work: OCR, field extraction, classification, and structured export from PDFs and scans. Low-confidence fields route to a human-in-the-loop verification step, keeping accuracy high while eliminating the manual data entry that eats hours.