ML Engineer — Video Understanding & Edge AI · San Francisco

Teaching systems to notice what changes.

I build vision‑language models that watch streaming video, hold a compact memory of what's normal, and reason deeply only at the moments that break the pattern. Work spans distributed pretraining, synthetic‑to‑real data pipelines, transformer acceleration, and edge deployment — from first pixel to production verdict.

Illustrative — perception stream vs. episodic memory trigger
Anomaly → memory recall
Research contribution — independent, 2026

Edge‑native action triage

From raw footage to a verdict, on‑device
Deployed
Weakly‑supervised · memory‑augmented · Jetson‑class inference

The project started as a data problem before it was a modeling one. I defined the action taxonomy, recording protocols, camera‑viewpoint specifications, and validation procedures behind a proprietary surveillance‑action corpus, then built the compact model that runs on it: 35,000+ verified clips and 137,000+ pose sequences after geometric augmentation, feeding a model small enough for an edge box.

Data engineering

Multi‑viewpoint capture, cleaning, scene‑level partitioning, controlled augmentation, and a strict held‑out split so no scene ever leaks between train and test.

Model design

View‑aware kinematic features feeding a FiLM‑conditioned Temporal Convolutional Network — 824K parameters, sized for Jetson‑class edge hardware.

Result

0.9661 weighted F1 across 27,511 sequences from scenes the model had never seen — evidence that data‑centric work plus a small model can beat a bigger one under an edge budget.

35K+
Raw clips
137K+
Augmented sequences
824K
Parameters
27.5K
Held‑out test
Availability

What's shareable

The underlying surveillance‑action dataset stays private — it was built inside a commercial research effort and isn't publicly distributed. A technical artifact package is available on request for research, academic, or hiring evaluation, covering:

  • Model architecture and feature‑engineering documentation
  • Dataset statistics, class distributions, and validation protocol
  • Training and inference benchmarks, plus a real‑time inference demo
  • Example pose‑sequence samples (non‑identifiable)

Request a copy: ayasin@shrinkhaltai.com

Instrumentation

Technical capabilities

Machine learning

  • PyTorch · TensorFlow
  • Transformers · VLMs · FlashAttention
  • Distributed training (DDP, FSDP)
  • Quantization · TensorRT · ONNX
  • RLHF / RLAIF
  • Self‑supervised learning

Cloud & systems

  • AWS — ECS, Lambda, API GW, SQS, DynamoDB, S3, CloudWatch, IAM
  • Docker · Kubernetes
  • CI/CD · monitoring · profiling
  • W&B · TensorBoard

Languages

  • Python · C++ · C# / .NET
  • SQL — MySQL, SQL Server
  • JavaScript · Bash

Edge & embedded

  • NVIDIA Jetson (Orin / Nano)
  • TensorRT · ONNX Runtime
  • Raspberry Pi · Arduino
  • LiDAR · embedded control
Chronology

Experience

2024 — present

Founding Machine Learning Engineer

ShrinkHalt AI · San Francisco, on‑site
  • Architected distributed training infrastructure for sensor‑fusion vision‑language models across NVIDIA GPU and AWS environments, with fault‑tolerant, scalable data pipelines.
  • Built distributed pretraining using data, tensor, and pipeline parallelism with optimized sharding and high‑throughput multimodal loading.
  • Accelerated transformer training with FlashAttention, mixed precision, gradient checkpointing, and memory‑efficient attention kernels.
  • Stood up experiment tracking and profiling with W&B, TensorBoard, PyTorch Profiler, and NVIDIA Nsight.
  • Deployed quantized multimodal models via TensorRT and ONNX Runtime, tuned for edge latency and throughput.
  • Developed weakly‑supervised video architectures using temporal attention, episodic memory, and long‑horizon behavioral modeling.
  • Built a generative data flywheel — automated curation, hard‑negative mining, pseudo‑labeling, continual representation learning.
  • Designed synthetic supervision pipelines for synthetic‑to‑real adaptation with artifact‑aware training.
  • Shipped end‑to‑end MLOps across ingestion, pretraining, SFT, RLHF/RLAIF, evaluation, and deployment.
PyTorch · Transformers · FlashAttention · TensorRT · AWS (ECS, Lambda) · Docker · W&B
2020 — 2023

Freelance Software Engineer

Freelancer.com · Dubai, hybrid
  • Delivered full‑stack systems for international clients across logistics, commerce, and enterprise workflow.
  • Built backend services in C# / ASP.NET Core with async processing and service‑oriented architecture.
  • Designed relational schemas and query‑optimized reporting pipelines on SQL Server and MySQL.
  • Ran cloud‑native workloads on AWS — ECS Fargate, Lambda, API Gateway, DynamoDB, S3, SQS, IAM.
C# · .NET Core · AWS (Lambda, ECS, S3) · Docker · SQL Server
2017 — 2020

Software Programmer & Data Engineer

Sungate Tyres Trading · Dubai, on‑site
  • Built backend and business‑critical applications in C#/.NET, with performance‑sensitive modules in C++.
  • Designed microservice architectures for inventory and operational automation.
  • Ran CI/CD, automated testing, and release management.
  • Tuned SQL Server workloads — indexing, query plans, stored procedures.
C# · C++ · SQL Server · Redis · Docker · Minikube · CI/CD
2014 — 2017

Software Programmer & IT Support

Softpro Network · Addis Ababa, on‑site
  • Developed and maintained applications supporting enterprise operations.
  • Translated stakeholder requirements into technical specs and documentation.
  • Built AJAX‑driven web apps in JavaScript, jQuery, JSON, HTML, and CSS.
C# · ASP.NET · jQuery · SQL Server · Windows Server
Foundations

Education

GISMA University of Applied Sciences

Potsdam, Brandenburg, Germany

MSc, Artificial Intelligence & Data Science

Research project: an autonomous mobile robot for beverage service, combining LiDAR navigation, YOLOv8 gesture recognition, Raspberry Pi 4, Arduino (ATmega328), and embedded vision.

Addis Ababa University

Addis Ababa, Ethiopia

BSc, Information Systems

Focus on software engineering, database systems, enterprise application development, and systems analysis.