Taalas HC1: An LLM Hardwired in Silicon
How chatjimmy.ai runs Llama 3.1 8B at ~15,600 tokens/s with the weights built into the chip.
Saved in this browser. Press t to cycle.
$ whoami
Chief Software Engineer, Samsung R&D · product search in 55+ markets · ACM-ICPC
I build search engines: query understanding, retrieval, ranking, and LLM reranking on top. Off the clock I read philosophy, listen to a lot of music, and play with algorithms.
Product search at Samsung, live in 55+ markets. A query goes in and ranked products come out. My team owns everything in between.
Normalise, correct, classify intent, expand. Work out what the person actually meant.
Lexical and semantic, fused. BM25 inverted indexes next to k-NN over HNSW graphs.
Ordering candidates by relevance with ML models and signals, tuned by measurement.
GenAI and LLMs over the top of the list, where precision matters most.
Spring and FastAPI microservices on AWS, with Redis caching and logging APIs. It has to stay up, globally.
I lead a search engineering team that owns the backend microservices, search and logging APIs, and relevance end to end. The platform started in one market and now runs in more than 55, and it has stayed highly available while search quality kept improving.
Before that I spent a long time on algorithms: ACM-ICPC, competitive programming, and 300+ hours teaching data structures.
$ git log --graph --oneline
Chief Software Engineer · Samsung R&D Institute India, Bangalore
BITS Pilani · B.E. Computer Science & M.Sc. MathematicsNotes, problem sets, and interactive experiments.
How chatjimmy.ai runs Llama 3.1 8B at ~15,600 tokens/s with the weights built into the chip.
The layered graph behind vector databases: how search and inserts work, M and ef, a 1M-vector benchmark.
Monte Carlo simulation, hands-on: throw darts for π, fill birthday rooms, test a project deadline.
Eight agent coordination patterns animated, plus replayable LangGraph and CrewAI runs.
How OpenAI watermarks ChatGPT text with optimal transport and an entropy budget, and how it's detected.
Rotating circles summing to draw any shape — a visual proof of the DFT.
Upload any image and transform it into pixel art with character overlays.
Space-efficient probabilistic data structures; implementation and false-positive analysis.
Interactive particle simulation with bouncing dots.
Programming language concepts and techniques.
Multiple approaches: triplets A[i] < B[j] < C[k].
Multiply polynomials in O(n log n); applications for large number multiplication.
Concepts and best practices for Elasticsearch.
Z-Algorithm, Trie, Suffix Array, Aho–Corasick, Hashing.
Curated list of 20 problems, Easy to Hard.
Problems using the DSU data structure.
Concepts and problem-solving techniques.
What we do, and why we do it.
He who has a why to live can bear almost any how.
Nietzsche
Psychology, consciousness, and the study of the human mind. What we do, and why we do it.
Without music, life would be a mistake. Sonic landscapes I keep coming back to.
listen →Learning Korean, from Hangul to grammar. Hangul is an algorithm too: jamo compose into syllable blocks.
my notes →From Samsung and elsewhere. Click a year to see the certificate.