Debanjan Basu — ML Systems Engineer

LLM agent observability, distributed pipelines, neural network compression research. Source for d3banjan.github.io.

Experiment artifacts (external)

Research artifacts live in their own repos and microsites, not in this one.

• Compression Falsification Ladder — methodology and rungs: compression-ladder-paper.pages.dev
• KV-Cache Quantization in MoE — moment-ratio framework, Lean 4 proofs: moe-gauge-paper.pages.dev
• LoRA-DPO Stable Rank — srank geometry, Lean-verified invariants, adapter checkpoints (~1.9 GB on HuggingFace): d3banjan.github.io/lazy-rudder-paper
• Symmetry Survey — algebraic invariants for compression: symmetry-survey-paper.pages.dev

Hardware used in published runs: RTX 3060 12GB · CUDA 12.x · Linux

Engineering surface

• LLM agent observability with Arize Phoenix — trajectory analysis for production agent debugging. Write-up: /series/methodology-transition/
• Distributed pipelines — Dask, Celery; 5–10× speedup on internal data workloads
• Agent security — type-stub gates for ML deserialization: /series/falcon-security/

Research methodology

Series: /series/methodology-transition/ External companion: compression-ladder-paper.pages.dev

Site dev

Built with Astro. Requires Bun.

bun install
bun run dev      # localhost:4321
bun run build    # ./dist/
bun run preview

CV is compiled from src/cv/cv.typ via Typst and served at /cv.pdf.