The systems behind the résumé.

Tasuki

Phase 1 of 7 done · proven locally

The delivery side of a business, as software you run in your own cloud. Your system creates a delivery over the API; Tasuki quotes a price and an ETA, dispatches the nearest courier on a real city road network, and streams the courier's position live to the customer's phone. It is sized for 1.5 million orders a month in one city and built to production standards: idempotent orders, a Kafka-backed location pipeline, written SLOs, load tests and game days.

orders/h at full design scale, on one laptop

18,144

GPS points/s, with 0 failed orders

641

p95 for POST /v1/orders under that load

20 ms

game-day plans; the first one run, with a postmortem

12

—The company is fictional and the merchants, couriers and customers are simulated. The AWS platform (Terraform, EKS) and the GitOps layer are written and statically checked, but not yet run on a cluster; that is deliberately the last phase.

Zero-retention AI KYC

Code private · architecture and API docs public

Identity verification an app can add with a few API calls: national ID card, selfie, proof of address and source of funds. Every document is quality-checked, read by open-weight vision models, cross-checked against the others, and passed to a compliance officer; the model never makes the decision, and nothing is kept.

two vision models: Qwen3-VL-8B and PaddleOCR-VL on one L40S

1 GPU

Terraform modules, content-addressed endpoint configs

12

API calls, each documented with a cURL example

11

GPU modes: always-on, scheduled, on-demand, off

4

—Client system: the code is private and is not linked. The architecture page, API docs and the write-up are public.

Private LLM Chat

Public · docs site with five guides

Your own ChatGPT-style app, on your own GPU. Open-weight models run with vLLM on Amazon SageMaker, several per GPU, with weights loaded from S3, an optional no-internet VPC, and a GPU that switches itself off when idle. The Django and React app streams chat and handles documents, images and voice.

open-weight model families: Qwen, Llama, Gemma, DeepSeek

4

GPUs running when idle: the endpoint scales to zero

→ 0

guides, from a laptop to the full AWS stack

5

Terraform stack: ALB, CDN, ECS, SageMaker, database, CI/CD

1

—A public open-source project with a docs site. No production users or traffic are claimed; the Google Cloud Run port is early.