hi, i'm himanshu!
machine learning engineer @ PrudentBit
stuck in the constant loop of learning new things and building cool stuff.
machine learning engineer @ PrudentBit
stuck in the constant loop of learning new things and building cool stuff.
currently, i'm a machine learning engineer at PrudentBit, where i build secure document intelligence systems — PII detection & redaction, multilingual NER pipelines, and the hybrid ML + rule-based stack behind them.
previously, i was a researcher at Lossfunk working on reducing compounding rollout error and its induced distributional shift, and studying RL generalizability on benchmarks such as Atari-100K.
i love deep learning, LLMs and world models, and am always looking to build cool projects on these — so ping me on any of the links below if you'd like to collaborate on something!
things i'm spending my time on:
architecting a secure document intelligence system for PII detection & redaction at enterprise scale — multilingual NER across structured and unstructured data (10k+ docs/day), pairing transformer models with rule-based validation for a ~10% F1 bump.
research on model-based RL for better generalization and lower sample complexity — prototyping world models for high-dimensional environments and chasing more stable long-horizon predictions.
built core modules of Immunefiles-PII, a production system for sensitive data detection, and helped ship scalable pipelines for real-world document workloads while tightening the precision-recall tradeoff on noisy data.
playing MsPacMan with RL + VQ-VAE world models
end-to-end pipeline for learning compact latent policies on Ms Pac-Man: a VQ-VAE compresses raw frames into a discrete latent space, DQN agents (with/without PER) learn directly on those latents, plus value / world-model / action-mapping nets for model-based planning.
sushi-GGUF
a minimalist framework for GGUF quantization of SDXL models — quantize Stable Diffusion to precision levels like Q4KS, Q5KS and Q8_0 using llama.cpp binaries.
paper implementations
a repo collecting several research paper implementations i've worked through — mostly reproductions to actually understand the ideas end to end.
i don't write much — most of what i become fascinated by, i share on my twitter. the longer pieces live here: