Finance meets technology.

“Khmer Netra Hub — a portal for everything in your infrastructure, one hub at a time.”

Self-hosted infrastructure operations behind one login, one theme, and one audit trail. A single Nuxt application presents a portal shell — launcher, sidebar, permissions, notifications, audit — in front of a growing suite of operations subsystems. Each subsystem is a Nuxt layer that merges into one in-process build: no micro-frontends, no iframes, no module-federation remotes. One image, one deploy.

• Six operations apps ship as layers — Docker Swarm console, Monitoring (a clean-room LibreNMS equivalent), Work (ClickUp equivalent), Privileged Access, IP Management, and Database Manager.
• Database per module on PostgreSQL + TimescaleDB with raw parameterised SQL and no ORM; disabling an app stops its work but keeps its data.
• Encrypted at rest end to end — LDAP binds, OIDC secrets, registry auth, SNMP credentials and the PAM vault under AES-256-GCM, with envelope encryption and a versioned master key.
• Scales horizontally with no broker — the database itself coordinates through advisory locks, lease-based job queues and LISTEN/NOTIFY. No Redis, no message bus.

“Enterprise AI call center between the Cambodia Telco SIP trunk and the 3CX PBX agent pool.”

Five separated services plus a shared scenario-schema library. A visual studio edits call scenarios as graphs; a worker executes those graphs per call against a pluggable ASR/LLM/TTS provider; a voice gateway drives FreeSWITCH over the Event Socket and hands live calls to 3CX agents. Every service is self-contained — own build, tests, Dockerfile — and liftable into its own repository.

• 844 automated checks, 0 failures in the 2026-08-09 QA campaign, with totals regenerated from the gates' own JUnit reports rather than hand-written.
• A real cross-service end-to-end suite boots the actual API, worker and gateway against a real PostgreSQL 18 — and found two defects every project's own green suite had missed.
• Two Docker Swarm stacks so the telephony half can live on a DMZ host: exactly three ports cross the boundary and every credential is a Swarm secret, never an environment variable.
• Per-tenant row-level security in PostgreSQL 18 under a three-role authorisation model that combines as a union.

Carrier onboarding, 3CX trunk provisioning and TLS/SRTP hardening are external steps the repository documents but cannot self-certify. The bundled AI provider is a deterministic simulator; real providers plug into the worker's SPI.

“Low-cost, camera-based driver drowsiness detection designed for retrofit into older vehicles.”

Risk is measured from eyelid closure over time, not from a whole-face “does this look drowsy” classifier — because a face classifier trained on DDD learned to recognise drivers rather than drowsiness. Eye closure is the mechanism itself, and it is cheaper on device: the model input is a 32×32 eye patch.

• A seven-wire build on an ESP32-S3-WROOM-1 N16R8 with an OV3660 camera and an I2S class-D amplifier. No display — the board serves its own Wi-Fi access point and a live MJPEG preview.
• 11.3k parameters, ~0.9 ms per frame for both eyes: a tiny grayscale CNN INT8-quantised to .espdl, realistic for an ESP32-S3.
• Four fused cues, not one — PERCLOS (0.55), long and slow blinks (0.20), yawning (0.15) and head nodding (0.10), each measured against a rolling per-driver baseline so face shape and camera angle cancel out.
• Sneezes are detected in order to be ignored: a sneeze slams the eyes shut for about a second, which a naive detector scores as a microsleep. Recognising it suppresses the false alert.

Research prototype, not a certified automotive safety device. Splits are subject-independent by principle, so no driver — and no adjacent frame of one clip — leaks across train and test.

“Windows desktop application for Claude organization owners — adoption monitoring, Weighted Usage Index, low-usage alerts and management reports from official Anthropic APIs.”

A small studio tool — a native Windows application shipped under PLXY — that answers the question every Claude rollout eventually gets asked: who is actually using this? It reads only officially supported Anthropic Admin APIs and official CSV exports — no scraping, no browser automation, strictly read-only — and turns them into dashboards, alerts and reports a manager can defend line by line in front of a finance committee.

• A number you can defend, not one that looks good — the Weighted Usage Index scores each account against configurable targets, weights and contribution caps, and shows the formula, policy version and per-metric breakdown in the UI and in every export.
• Missing data is missing, never zero — every data category is probed and shown as Available or Unavailable with the reason, so nobody is flagged a low performer because of an API gap in their plan.
• Alerting with governance — thresholds per group, new-user grace periods, duplicate suppression, and a fully audited Alert Exemptions register. Monitoring only: accounts are never modified and conversation content is never touched.
• Seven export formats from one snapshot — PDF, XLSX (11 worksheets), DOCX, PPTX, CSV, HTML and JSON, so exported numbers exactly match the screen. API keys live in Windows Credential Manager, never in a file.

Proprietary software — the source repository is private, so unlike the other headline projects there is no public code link. The documentation, download page and SHA-256 checksums are public, and Demo Mode runs the whole application on an anonymized synthetic organization, no credentials needed.