~/projects
Projects
Things I've built and operated. Expand any entry for the full story.
End-to-end tool for collecting, annotating, and exporting person-attribute datasets — built to make annotation work dramatically easier.
+ story− story
// problem
Annotating person-attribute datasets from scratch is slow and error-prone, and near-duplicate images leaking across train/val/test splits silently inflates validation scores.
// what i built
A full pipeline tool — upload, batch, crop, assign, pre-label, annotate, split, export — with model-assisted pre-labeling (PULC person-attribute model + YOLO person crops), multi-user annotation with per-user leaderboards and resume-where-you-left-off, and bit-level correction tracking of who changed what.
// technical decisions
Model pre-labels so annotators correct rather than label from zero; group-aware splitting using 64-bit perceptual hashing so near-duplicates can't leak across splits; atomic per-row upserts with SQLite WAL for concurrency-safe annotation writes.
// challenges
Keeping multi-user annotation correct and concurrent — coverage tracking, overlap warnings, safe writes — on a single-machine SQLite deployment.
Python · FastAPI · React · SQLite · YOLO · PaddleClas
Visibel.ai
Backend EngineeringComputer Vision & AIproductionFullstack + ML platform for AI service registration, execution management, and monitoring — plus research and training of models for various tasks.
+ story− story
// problem
The platform needed to run AI services reliably in production while also supporting ongoing model research — and do both on edge machines with limited storage, memory, and processing headroom.
// what i built
Both frontend and backend for a platform handling AI service registration, execution management, monitoring, and operational reliability, alongside research and training of models for various tasks.
// challenges
Keeping the platform, models, and services correct and efficient on edge hardware with real resource constraints — storage, memory, and processing — without sacrificing performance or reliability.
Backend · Frontend · Model training · Edge deployment
Vessel Management System
Backend EngineeringConnected SystemsproductionLed architecture and backend for a vessel management platform — real-time fleet telemetry, satellite integration, and operational reporting.
+ story− story
// problem
The project wasn't about writing new features from scratch — it was ensuring critical workflows from a legacy system kept working correctly in a new one, while improving them. That meant reverse-engineering how the old system actually behaved in production, not just how it was supposed to behave.
// what i built
As Technical Lead and Backend Engineer, I led a 2-person engineering team, designed the system architecture, and implemented real-time fleet telemetry processing and operational reporting — including vessel tracking with satellite integration: a satellite comms client, a background scheduler polling a mailbox and generating “laporan-inmarsat” reports, and a geotrack scheduler for position sync.
// technical decisions
Rather than a straight migration, we reverse-engineered legacy behaviors and hidden business rules first, then rebuilt them in a cleaner, more maintainable architecture — improving weak points without breaking what operators already relied on.
// challenges
Preserving operational continuity while replacing the legacy system underneath it, without disrupting established user expectations.
// result
Delivered in roughly 9 months, supporting a fleet of 100+ vessels (client base may exceed that).
Python · FastAPI · SQL · RBAC · Telemetry · Satellite comms
Data Center Infrastructure Management (DCIM)
Backend EngineeringConnected SystemsComputer Vision & AIproductionBackend engineering for a DCIM platform monitoring temperature, electricity, and visitor count, plus AI Vision detection — on a 4-person team.
+ story− story
// problem
A data center needed infrastructure monitoring that could ingest and aggregate device data across systems — environmental sensors, visitor counting, and camera-based detection — and expose it through reliable APIs.
// what i built
As part of a 4-person team, backend services for device data ingestion, aggregation, and monitoring APIs covering temperature, electricity, and visitor-count sensors, plus integration of AI Vision monitoring with 3 custom-trained models (fire, trash, and rodent detection).
// technical decisions
Deployed and wired DCIM together with a VMS and the AI monitoring platform rather than building each in isolation, so device and vision data shared a common ingestion/monitoring layer.
Backend · APIs · Data aggregation · Deployment · Computer vision
Visitor Management System
Backend EngineeringConnected SystemsproductionDiagnosed and fixed a production scaling incident on a visitor/access management platform, then followed up with dashboard and authorization features.
+ story− story
// problem
In production, devices sending reports simultaneously over an extended period caused responsiveness problems — a single service handled both high-volume device ingestion and user-facing frontend/API traffic at once.
// what i built
Split the service in two — one dedicated to receiving device reports, another to frontend/API operations. With the backend isolated and stable, followed up with a refreshed dashboard, an overhauled person page, a devices view, device grouping, and a feature authorizing specific people on specific devices.
// technical decisions
Isolated workloads by responsibility rather than patching the shared service, so each traffic pattern could scale independently.
// challenges
Ruling out device behavior, database write performance, RAM/CPU usage, and local network stability one by one before finding the real bottleneck: one service doing two very different jobs.
// result
Responsiveness improved significantly and both services now scale independently — the dashboard and authorization features that followed were built on a healthier backend.
Backend · Frontend · Device APIs · Authorization