SV
GH IN EM
Sashit Vijay

Useful systems with full-stack craft

I build full-stack products, machine learning systems, and computer vision tools with a focus on speed, clarity, and interfaces that feel calm enough to trust.

See the work or walk the 3D house →
CorpusForge v3 · 54M params
dataset: OpenWebText · 50M tokens layers: 8 · heads: 8 · d_model: 512 · ctx: 256 $ python train.py --model v3 --steps 5000 step 1000 · loss 5.82 step 3000 · loss 5.38 step 5000 · train_loss 5.18 val_loss 5.22 ✓ checkpoint saved · serving via FastAPI █

01 · 一

GPT decoder built from scratch

54M-parameter decoder-only transformer trained on 50M tokens of OpenWebText with no pretrained weights. 8 layers, 8 attention heads, 512-dimensional embeddings, Flash Attention, cosine LR with 200-step warmup, and AdamW with gradient clipping. Inference served via FastAPI.

PyTorchFastAPItiktokenFlash Attention
View project

02 · 二

Computer vision for match analysis

Processes Bundesliga footage through a YOLOv8 detection pipeline, KMeans shirt-colour team segmentation, optical-flow camera motion compensation, and perspective transform — producing real-world speed and distance per player from raw video frames.

PythonYOLOv8OpenCVKMeans
View project
Onsight Pipeline ep 382 – 777
NER SpaCy · entity extraction
Graph NetworkX · co-occurrence
Classify DeBERTa · ability types
Theme zero-shot · DeBERTa-large
Chat Llama 3.1-8B · LoRA

03 · 三

Five-module NLP pipeline

End-to-end LLM/NLP system analysing 395 episodes of One Piece dialogue. SpaCy NER for entity extraction, NetworkX character co-occurrence graph, fine-tuned DeBERTa ability classifier, zero-shot episode theme analysis with DeBERTa-large, and a LoRA-tuned Llama 3.1-8B chatbot — all unified in a single Gradio dashboard.

Llama 3.1DeBERTaSpaCyGradio
View project
SyncLink NFS · 3-tier distributed FS
clients: 50 concurrent · protocol: binary 9-byte header $ nm --storage SS_Root_0,SS_Root_1 network manager bound · routing ready $ client write /docs/report.pdf 2.1MB → SS_Root_1 · acked in 11ms $ client copy /docs/report.pdf /backup/ ✓ cross-server copy complete █

04 · 四

Distributed file system in C

Network File System written in C supporting 50 concurrent clients across a three-tier architecture — Clients, Network Manager, and Storage Servers. Binary protocol with a 9-byte header (1-byte type + 8-byte length). Full CRUD, cross-server copy, metadata retrieval, and backup management at the Network Manager layer.

CPOSIX SocketsBinary ProtocolNFS
View project

Experience · 経歴

role_01.ts 
const role_01 = {
  title:   "LLM Evaluator",
  org:     "Deel Inc.",
  since:   "2026",
  focus:   "test-authoring benchmarks",
  targets: { frontier: "80%+", baseline: "25–50%" },
};

2026 → present

01 · 一

LLM test-authoring evaluator

Builds reproducible LLM evaluation benchmarks for test-authoring agents using rubric-calibrated scoring, Dockerized Harbor sandboxes, hidden RewardKit rubrics, oracle diffs, and regression-guarded test scenarios — targeting 80%+ for frontier models and 25–50% for baselines.

LLM EvaluationDockerRewardKitBenchmarking
role_02.ts 
const role_02 = {
  title:  "Fullstack Developer Co-op",
  org:    "Bytewerx LLC",
  period: "Aug 2024 – May 2025",
  api:    "Tnect Validation API",
  built:  ["RBAC", "rate-limiting", "Redis"],
};

Aug 2024 → May 2025

02 · 二

API platform engineering

Built a secure API management platform with Angular, ASP.NET Core, TypeScript, and Redis — scoped key generation, role-based access control, Redis-backed rate limiting, TypeScript-typed service layers, and a real-time analytics dashboard for the Tnect Validation API.

AngularASP.NET CoreTypeScriptRedisRBAC
role_03.ts 
const role_03 = {
  title:    "Research Assistant",
  lab:      "LLEAS Lab · ASU",
  period:   "Oct 2024 – Jul 2025",
  platform: "Unitree Go2",
  latency:  "<200ms",
};

Oct 2024 → Jul 2025

03 · 三

Robotics & AI research

Developed a Unitree Go2 robotic dog assistant with real-time diffusion-based navigation and LLM-powered command processing. Integrated obstacle avoidance achieving sub-200ms response latency — aimed at improving independence for visually impaired users.

PythonDiffusion ModelsLLMsROS2PyTorch

Built with clarity, taught with patience.

I am a computer science graduate from Arizona State University with a 4.0 GPA, Summa Cum Laude honors, and a mix of product, research, teaching, and systems experience.

My work tends to sit where useful engineering meets clear communication: backend systems that hold up, ML workflows that explain what they are doing, and frontends that do not make users fight the interface.

Send me a message

Tool garden · 道具

Python Java C++ C React Node.js Django Flask PyTorch OpenCV MongoDB MySQL NumPy Pandas scikit-learn MATLAB

Contact · 連絡

Have a project or idea that needs a builder?

sashitvijay@gmail.com Resume

Interactive · Three.js · 3D walkthrough

Enter Sakura House

A real-time 3D walkthrough of a lakeside Japanese home — explore every room at your own pace.