Turning research-grade NLP into systems people actually use.
Rakshit Makan is a Generative AI Specialist based in Toronto with 8+ years across AI, NLP, and Deep Learning. He works on large language models, retrieval-augmented generation, and transformer fine-tuning — the full path from research and experimentation to production systems.
His master's thesis at Dalhousie University built a human-in-the-loop system for classifying multi-page administrative documents, and his work since spans document clustering with BERT, modular LLM training and inference pipelines, and agentic AI tooling. He currently builds at Behavox, and cares about clean ML pipelines, well-scoped models, and tooling that removes friction.
Things I've built
Human-in-the-Loop Document Classification
The reference implementation of my master's thesis. A human-in-the-loop system for classifying multi-page scanned administrative documents: a two-step Sentence-BERT classifier (Page Importance → Main Page) where a domain expert validates and annotates, and the model retrains incrementally — no ML engineer in the loop. Built with Flask, Redis, and scikit-learn.
Document Clustering with BERT
Unsupervised topic clustering using BERT & SBERT embeddings, PCA dimensionality reduction, and K-Means. SBERT gave the cleanest topic separation.
LLM Training & Inference
A modular Python application for training and running language models across both Hugging Face and ONNX formats.
Claude Job Hunter
End-to-end job-application automation as a Claude Code plugin — discover, filter, tailor resumes, and auto-apply.
AI Marketing Automation
LLM-driven content generation and campaign workflows, automating the repetitive parts of marketing.
What I work with
Recent essays
I Turned My Job Search Into a Pipeline
Building the claude-job-hunter plugin — 60 applications in one night, 7 recruiter calls, and what it taught me about agents.
Anatomy of a Claude Code Plugin
The architecture under the hood: skills, hooks, MCP, and a files-as-state pipeline — with almost no traditional code.