rec ● live
fov 1:1 · iso 2026

const focus = ""

Software Engineer building AI systems that survive contact with production.

7+ yrs
shipping software
10min → 1min
pipeline optimization
97%
cv model accuracy
Toronto, Canada
based in

01.Signal, not noise

I build AI systems that actually ship — the unglamorous, load-bearing part of machine learning where models meet Docker containers, message queues, and production traffic.

My strongest work sits at the intersection of ML and software engineering: turning models, image-processing pipelines, APIs, databases, and deployment infrastructure into systems that run reliably at scale. Most recently that means an AI-powered PCB inspection platform — board scanning, component analysis, multi-board comparison, real-time inference.

I also evaluate frontier-model outputs for code generation, reasoning quality, and tool use — which has sharpened my sense for the difference between code that looks correct and code that is actually reliable.

// education.log
Lambton College, Toronto
Post-Graduate — Artificial Intelligence & Machine Learning
2024
Kathmandu University
B.Sc. Computer Science
2019

02.Deployment history

Cybord AI

2025 — Present
Software Engineer · Remote
  • Engineered the microservice backend for a PCB inspection POC — board scanning, component analysis, multi-board comparison — accepted by SpaceX and now in production.
  • Built a production desktop app (Electron + FastAPI) integrating proprietary AI models with real-time inference.
  • Cut pipeline processing time from ~10 minutes to ~1 minute with task-graph parallelization; reduced model inference time by 60%.
  • Fine-tuned computer vision models to 95% production accuracy and automated visual data extraction, saving 200+ engineering hours monthly.

DataAnnotation

2024 — Present
AI Training Specialist (Freelance) · Remote
  • Evaluate frontier-model outputs: code generation, debugging, reasoning quality, and instruction following.
  • Identify failure modes in multi-step reasoning, tool-use workflows, and agentic coding tasks.
  • Surface the subtle issues automated benchmarks miss — plausible-but-wrong code and unsafe implementation choices.

Vanilla Transtechnor

2021 — 2022
Software Engineer · Kathmandu, Nepal
  • Built scalable web applications serving 5,000+ daily active users with Django REST Framework and React.
  • Standardized auth across 12+ microservices with custom middleware, permission classes, and mixins.
  • Automated deployment pipelines, cutting deployment time by 40%.

Eightsquare

2019 — 2021
Software Engineer · Kathmandu, Nepal
  • Pioneered a mobile SDK for Malaysian identity-document OCR (IDs, passports, visas), 80% faster through OpenCV optimization in C++/JNI.
  • Shipped face recognition + liveness detection SDK — 97% accuracy, 10,000+ daily authentication requests.
  • Architected a multi-tenant push-notification SaaS for 50+ enterprise clients at 99.8% uptime.
  • Built an employee management system serving 60,000+ employees with role-based access control.

03.Systems in the wild

[computer_vision]conf 0.99

PCB Inspection Platform

AI-powered board scanning, component analysis, OCR/marking extraction and multi-board comparison. The POC backend was accepted by SpaceX; now running in production.

PythonKafkagRPCFastAPIRedisOpenCV
10min → 1min
pipeline time
[ml_infrastructure]conf 0.97

Real-time Inference Desktop App

Production Electron + FastAPI desktop application integrating proprietary AI models with a multiprocessing inference pipeline for enterprise deployment.

ElectronFastAPIPyTorchMultiprocessing
-60%
inference time
[mobile_cv]conf 0.96

Identity Document OCR SDK

Android SDK for Malaysian ID cards, passports and visas — real-time image processing through OpenCV via JNI (C++), tuned for on-device speed.

AndroidC++JNIOpenCV
+80%
performance
[biometrics]conf 0.97

Face Recognition & Liveness SDK

Face recognition and liveness detection under varied real-world conditions, processing 10,000+ authentication requests daily.

PythonDlibTensorFlowOpenCV
97%
accuracy
[backend_saas]conf 0.95

Multi-tenant Push Platform

Push-notification SaaS supporting 50+ enterprise clients, plus an employee management system serving 60,000+ users with role-based access.

DjangoPostgreSQLRedisRabbitMQ
99.8%
uptime
[ai_evaluation]conf 0.94

Frontier Model Evaluation

Structured evaluation of frontier-model code generation and agentic tool use — comparative rankings and failure-mode analysis for RLHF pipelines.

LLMsCode ReviewAgentic Workflows
RLHF
feedback loop

04.Toolchain

// Machine Learning & Vision
PyTorchTensorFlowOpenCVComputer Visionscikit-learnpyvipsModel TuningNumPyPandas
// Distributed Systems & Backend
PythonasyncioFastAPIKafkagRPCProtobufRedisCeleryRabbitMQDjangoPydanticDependency Injection
// Infrastructure & Observability
DockerAWSS3OpenTelemetryPrometheusCI/CDuvMicroservices
// Data & Frontend
PostgreSQLSingleStoreSQLAlchemyMySQLDynamoDBReactNext.jsElectronTailwind CSS
PyTorchTensorFlowOpenCVComputer Visionscikit-learnpyvipsModel TuningNumPyPandasPythonasyncioFastAPIKafkagRPCProtobufRedisCeleryRabbitMQDjangoPydanticDependency InjectionDockerAWSS3OpenTelemetryPrometheusCI/CDuvMicroservicesPostgreSQLSingleStoreSQLAlchemyMySQLDynamoDBReactNext.jsElectronTailwind CSSPyTorchTensorFlowOpenCVComputer Visionscikit-learnpyvipsModel TuningNumPyPandasPythonasyncioFastAPIKafkagRPCProtobufRedisCeleryRabbitMQDjangoPydanticDependency InjectionDockerAWSS3OpenTelemetryPrometheusCI/CDuvMicroservicesPostgreSQLSingleStoreSQLAlchemyMySQLDynamoDBReactNext.jsElectronTailwind CSS

05.Open a connection

status: accepting_interesting_problems

Production ML, computer vision, AI infrastructure — if you're building something real, open a channel. No naked email address on this page; the terminal handles delivery and the scrapers starve.

$ scp resume.pdf ./local ↓
suyog@mailer:~ — tx
secure channel established. handshake ok.
3 fields required to open a connection.
[1/3] identify yourself: