# Kannav Sethi

> AI applications developer building LLM tools and workflows. Writing about agents, reliable software, and inference.

Kannav Sethi is an AI applications developer. Contact: kannavsethi02@gmail.com.

## Work

- Software Engineer, [Locate Alpha](https://locatealpha.com) (current)
- AI Engineer, [Acclaim Ability](https://www.acclaimability.com) (current)
- Product Builder, Seneca Polytechnic (previously)
- Software Dev Intern, Project Human City (previously)

## Research

- [Do Query-Agnostic Memory Structures Earn Their Budget? A 10M-Token Audit with Exact Read Limits](https://aclanthology.org/2026.realm-1.18/): Kannav Sethi, Amit Maraj. REALM Workshop at EMNLP 2026. Research with Amit Maraj on whether organizing an agent's memories improves its answers. We compared memory organization approaches with direct passage search under the same reading budget.

## Writing

- [Controlling Probabilistic Outcomes by Ensuring a Deterministic System](https://kannav.dev/blog/deterministic-systems-for-probabilistic-outcomes) (Jul 2026): I spend more time on X/Twitter than I’d care to admit, and over the past few weeks, one debate has kept resurfacing, whether or not to read AI-generated code.
- [Local Models, Unified Memory, and Why Quantization Matters](https://kannav.dev/blog/local-models-unified-memory-quantization) (May 2026): I have been exploring running large language models, specifically Qwen3.6 9B, alongside inference stacks like Ollama and llama.cpp-style runners, single sequence, no batching. I've also been digging…
- [A (not so big) Primer on GPU Internals](https://kannav.dev/blog/gpu-internals-primer) (Mar 2026): For the past couple of weeks, I've been exploring things that would be exciting to learn, be it how to structure workflows with different agentic harnesses, a couple of math stuff with classical MLs…
- [Floats and Precision in LLM Training](https://kannav.dev/blog/floats-and-precision-in-llm-training) (Feb 2026): A while ago, I was SFT-ing a Qwen 0.6B model on the UltraIntellect dataset to understand how the training loop behaves. While running training locally, I ran into a CUDA (OOM) issue, yup it happened…
- [Normalization Techniques](https://kannav.dev/blog/normalization-techniques) (Dec 2025): As part of me exploring different internal things about LLMs, this blog is about, if not the most, but one of the most important and misunderstood concepts in the field of LLMs, Normalization…
- [Rotary Positional Embeddings and its Math](https://kannav.dev/blog/rotary-positional-embeddings) (Dec 2025): I recently implemented a couple of different architectures for LLMs, its quite nice to see how different each architecture is from each other and how they solve problems stemming from the same core…
- [KV Cache, MQA, GQA and Attention!!!!](https://kannav.dev/blog/kv-cache-mqa-gqa-attention) (Nov 2025): I have been planning to write a blog post on a lot of things I have been reading/playing around with over the past couple of weeks. I explored and tinkered around a lot of core concepts related to…

## Projects

- [Inference Engineering Companion](https://inference.kannav.dev/start/reading-map/): something I built to help me understand inference: a reading companion to Philip Kiely's Inference Engineering, with original labs on local inference, KV caching, memory budgets, and benchmarking
- [X Bookmark Engine](https://github.com/Kannav02/x_bookmark_engine): a local bookmark CLI combining keyword and semantic search, with local models for embeddings and tagging
- [tiny-doc-cli](https://github.com/Kannav02/tiny-doc-cli): a TypeScript CLI that extracts readable content from PDF, DOCX, XLSX, and CSV files for agents and scripts
- [DialectMorph](https://github.com/Kannav02/DialectMorph): an AI-assisted CLI for translating code between Python, JavaScript, Java, and C++

## Links

- [Website](https://kannav.dev)
- [GitHub](https://github.com/Kannav02)
- [LinkedIn](https://www.linkedin.com/in/kannav-sethi-055313237/)
- [X](https://twitter.com/kannav02)