Software engineer. Stack-agnostic. I build applied AI systems end to end not demos, not notebooks, not "proof of concepts" that live permanently in a Notion doc. The full problem: architecture decisions, backend design, user experience, pipeline engineering, cost routing, retrieval accuracy, and making sure the thing keeps working once real users get their hands on it. The model is one component. The system around it is the actual work.

Currently deep in Vision Transformers after finishing a structured empirical comparison between CNNs and ViTs that might become a paper. Also building production LMS and document AI products where the actual engineering challenge is the system design around the model, not the model itself.

• DocuChat — Multi-document RAG with real-time passage highlighting and character-level source attribution. Next.js + Flask microservice + PHP CodeIgniter. Live →

• UpSkill AI — AI course platform that generates structured content and gates progression behind quiz checkpoints. Multi-provider LLM routing (Groq + Cerebras + Gemini). 50 subtopics in 24 seconds. Walkthrough →

• CNN vs ViT Study — Empirical comparison across CNN baselines, ResNet variants, and Vision Transformer (scratch + pretrained) on CIFAR-10. 95.4% with fine-tuned ViT. Working toward a paper.

Commits are the honest number. Stars are a marketing metric. The rest catches up when the products ship.

 

• Why most AI product demos are impressive and most AI products are not
• The gap between "it works in a notebook" and "it works at 3am under real user load"
• Building things I wish existed, not because the idea is unique, but because I want them to exist
• Reinventing the wheel on purpose to understand the architecture from scratch, then putting something new on top. The wheel isn't the goal. What you build after understanding it is
• Why ViTs beat CNNs at scale but CNNs still win on small datasets — and what inductive bias has to do with it
• System design decisions that cut LLM API costs without touching output quality (the UpSkill video pipeline is a good example of this)
• Character-level source attribution and how it changes the trust dynamic in RAG systems
• Stay curious. Learn new things. Pick up whatever the problem needs, not whatever you already know

• Agent architectures → what actually makes multi-step reasoning reliable
• CNN vs ViT paper → writing up findings, might publish
• Vision Transformer training run → in progress
• LLM inference internals → KV cache, speculative decoding
• Always something new on the side — that's the point

If you're building something where the hard parts are system design, backend architecture, or the AI layer that has to actually work in production or if you just want to talk about something interesting you're working on, reach out.