Aayush Garg
Aayush Garg
Forward deployed ML engineer at Applied Computing
I build and deploy AI systems, from first principles to production.
Right now that means building AI models and a reasoning layer for refineries and plants optimization at Applied Computing, from the first client call to production. Before that, I built image generation and enhancement pipelines at Jiffy and Flixstock, with some contract work on the side. Outside work, I tinker, rebuild language models from scratch and write about what I learn.
Now October 2026
Working on
AI models and a reasoning layer for refineries and plants
Applied Computing
Building
FlashAttention2 from first principles
building-from-scratch
Live
Updated every 15 minutes
Latest post
Kalman Filters and Smoothing for Time-Series Data
15 Jul 2026 · 16 min
Path
-
2015 to 2022
Research
Applying deep learning to subsurface imaging and inversion.
- TU Delft 2015 PhD Researcher JMI-res for high-resolution subsurface imaging, and deep learning super-resolution for seismic data.
- Shell 2019 Research Geophysicist Deep learning across seismic research: denoising, uncertainty quantification and learned Hessians for wave-equation optimization.
-
2022 to 2026
ML development and engineering
Building, deploying and running generative image pipelines in production.
- Flixstock 2022 Software Development Engineer III Stable Diffusion pipelines for photo-shoot backgrounds, identity transfer and fashion video, steered by custom ControlNets.
- Jiffy 2024 Senior ML Engineer The image enhancement pipeline behind a $100M+ DTF printing business, and Smart Redraw, real-time multi-agent image editing, on a GPU inference stack tuned for latency and cost.
-
2026 onward Now
Applied ML
Production AI for refineries and plants.
- Applied Computing 2026 Forward Deployed ML Engineer Building and deploying AI models and a reasoning layer for refineries and plants, from the first client call to production.
Writing All writing
- Time Series 16 min
- Tools & Infra 9 min
- Local LLMs 16 min
- Local LLMs 6 min
- Local LLMs 11 min
Built All work
-
building-from-scratch
Language models from first principles: GPT-2, a BPE tokenizer with a Rust encoder, KV cache, SFT and GRPO.
-
India Weather
Weather and air quality across Indian cities on one map, with 24-hour history.
-
Browser tools
Image, JSON, YAML, LaTeX and diff utilities that run entirely in your browser.