Catchify is an automated driver and vehicle verification platform designed for Bolt Georgia to replace manual document checks with continuous, fraud-resistant verification. The project won Bolt's internal Shark Tank competition and moved into active implementation.

The platform processes 5,000+ verifications monthly with 95% approval accuracy and reduces manual review volume by 60%. The web pipeline runs a live-capture camera flow with SMS dispatch, blocking pre-recorded uploads to reduce tampering and spoofing. Behind the capture layer, Catchify combines vehicle-damage detection, document OCR, and identity matching. Its four-layer anti-fraud architecture includes screen-replay analysis, camera sensor fingerprinting (PRNU), Error Level Analysis (ELA), and synthetic-media detection.

GhostPilot is an invisible desktop copilot designed for real-time interview assistance and meeting analysis. It operates as a local-first transparent overlay, processing user microphone input, meeting audio loopbacks, and screen capture in parallel.

I engineered independent, non-blocking IPC pipelines and capped memory consumption with bounded two-minute PCM ring buffers. Model streams are canceled through AbortController signals, while renderer tokens are batched with requestAnimationFrame to keep high-frequency output below a 16ms UI frame budget. The system supports multiple cloud model providers alongside a local whisper.cpp sidecar for offline speech transcription.

A centralized internal evaluation and quality assurance platform built for Bolt Georgia to replace fragmented third-party monitoring tools with automated dispute resolution and performance auditing.

Architected secure data access with role-based access control (RBAC) and row-level security (RLS). The platform reached 85% company-wide adoption and 100% adoption across four operational departments, standardized quality-evaluation metrics, saved €12,000 annually in software licensing, and reduced ticket-resolution cycles by 25%.

A production-style, multi-tenant RAG platform built with FastAPI, SQLAlchemy, PostgreSQL, Redis, Dramatiq, Alembic, and Swagger. It combines authenticated document ingestion with durable background extraction, deterministic normalization, structure-aware chunking, provenance, and owner-scoped persistence.

Dramatiq and Redis workers process documents through integrity verification and retry-safe indexing. The retrieval layer returns enterprise document results in under one second and reduced information-search time by 50%. The system retains short-lived JWT authentication, Argon2id password hashing, bounded retries, advisory locking, and atomic persistence across the ingestion and retrieval pipeline.

High-frequency data engineering pipelines designed to track, normalize, and analyze competitor pricing fluctuations and localized demand shifts across regional ride-hailing markets.

Automated end-to-end data ingestion across disparate public endpoints and geospatial feeds, processing more than 100,000 daily price observations. Built anomaly-detection filters and executive dashboards that helped operations teams calibrate incentive multipliers and surge thresholds.

A predictive machine learning pipeline designed to analyze operational disruption patterns across urban ride-hailing networks.

By processing over 148,000 historical ride bookings, I engineered features tracking spatial demand density, driver availability, and surge-pricing volatility. The classification models achieved 89% accuracy and a 0.96 AUC in forecasting cancellation risk. I translated the outputs into real-time operational dashboards and proactive driver reassignment, reducing cancellations by 12%.

Engineered municipal geospatial data pipelines and a routing API using Dijkstra's algorithm to model urban traffic topologies and infrastructure constraints. Dynamic graph pathfinding reduced fleet route latency by 12% across more than 500 simulated vehicles.