Ashutosh Ranjan

A 124M-parameter transformer trained end to end in raw PyTorch — RoPE, RMSNorm, SwiGLU, KV-cache attention and a byte-level BPE tokenizer, all hand-written across 6,000 lines. No transformers.AutoModel: every layer implemented and measured.

• 23.53 validation perplexity on 2.46B tokens of FineWeb-Edu — 20,000 steps, 27 h on a single RTX 2080 Ti at 25K tok/s.
• Cut peak memory 50.9% with gradient checkpointing, accumulation and fused attention, all gradient-identical; FSDP with CPU offload shards a 1.23B config to 2.04 GB/GPU.
• Sequence packing reclaimed the 89% of each fine-tuning step lost to padding, for 4.40x supervised throughput; DPO on HH-RLHF moved the win rate from 50% to 57.1%.
• Served from Hugging Face Spaces behind a streaming, KV-cached FastAPI endpoint at 80 tok/s, under 441 tests on GitHub Actions.

• PyTorch
• CUDA
• DDP / FSDP
• FastAPI
• HF Spaces

• Code
• Live demo

A Node.js control plane that schedules training onto real Kubernetes Pods: its own scheduler enforcing quota and placement (and printing every node it rejected), an executor driving Jobs from observed Pod status, and PostgreSQL state behind an append-only event log.

• Runs are dequeued with SKIP LOCKED, so concurrent workers never collide on the same job.
• Checkpoints upload to MinIO/S3 presigned and are verified against the bucket before any model version may use them.
• One PRODUCTION version per model, promoted transactionally, then served behind a weighted canary router.
• ResNet-18/CIFAR-10 proved the whole path: one epoch in 691 s, 65.59% top-1 re-evaluated from object store — SIGKILL the pod and it resumes weights, optimizer, LR schedule and data order.

• Node.js
• Kubernetes
• PostgreSQL
• MinIO / S3
• Docker
• Prometheus

• Code
• Live
• Demo

A streaming agentic loop exposing 12 tools — file edit/patch with diffs, shell with timeout and scrubbed env, grep/glob, web search — plus MCP servers over stdio and HTTP/SSE, driven by any OpenAI-compatible endpoint.

• Survives long sessions in a finite context window: history compacts at 80% capacity and stale tool output is pruned.
• A signature-based detector breaks loops of repeating failed tool calls instead of burning the budget.
• Six approval policies gate destructive shell and filesystem operations.
• Shipped to PyPI under 36 tests on GitHub Actions across Python 3.11 to 3.13.

• Python
• MCP
• asyncio
• GitHub Actions

• Code
• PyPI
• Demo