class IshworSubedi(MLEngineer):
role = "Machine Learning Engineer @oxura.dev"
based_in = "Kathmandu, Nepal (UTC+05:45)"
builds = ["computer vision pipelines", "LLM & agentic applications"]
stack = ["PyTorch", "OpenCV", "FastAPI", "Docker", "PostgreSQL"]
exploring = ["agent orchestration", "retrieval quality", "on-device inference"]
def working_on(self) -> str:
return "taking research-grade models to production-grade reliability"
def open_to(self) -> list[str]:
return ["collaboration", "applied AI/ML roles", "a good technical argument"]|
Detection, tracking and counting on live video. YOLO-family models, custom training, OCR and ANPR, deployed against real camera feeds instead of curated datasets. |
Retrieval pipelines, tool-using agents, and generative interfaces — built with prompt evaluation and failure handling as part of the design, not an afterthought. |
The unglamorous half: FastAPI services, containerised inference, versioned artefacts, monitoring, and latency budgets that hold under load. |
| Project | What it does | Why it was hard |
|---|---|---|
| Person Tracker & Counter | Real-time detection, tracking and persistent counting from a video feed | Keeping identity stable across occlusion and re-entry without inflating the count |
| Trinetra — ANPR | Nepali number-plate recognition plus traffic-light violation detection | Devanagari plates, motion blur, and night footage break off-the-shelf OCR |
| ICO-Gen | Text-prompt icon generation, end to end in TypeScript | Turning a generative model into a product with a usable interface |
| ML Algorithms | Notes, implementations and projects from the ground up | Written to be read — the reference I wish I'd had while learning |
| Languages |
|
| ML & Vision |
|
| LLM & Agents |
|
| Data |
|
| Serving & Ops |
|
“The magic you are looking for is in the work you are doing.”
