Research, stories, and insights on AI and Data Engineering.
Deep dives into neural networks, NLP, and AI systems
Pipelines, warehouses, GPU-accelerated query engines, and big-data systems
Illustrated novels that teach the modern AI stack as story, with audio narration in English and Hindi
With gratitude to every unsung hero across AI and Data — the engineers, researchers, and builders whose quiet work makes all of this real.
— P.S.
What is an LLM, intuitively?
Tokens & embeddings — text becomes math.
Vector spaces, similarity, and the moment they named the model.
Self-attention, and a stranger at the library.
Their toy model says a sentence neither of them was ready for.
Layers, residual streams — and the secret hours.
samkhya v1.0 ships an LLM-pluggable corrector backend for embedded analytical engines — DataFusion, DuckDB, Polars, Postgres, Iceberg, gpudb. Plug Claude, GPT-4o-mini, or local Ollama into the cardinality-estimation slot via a simple HTTP wire contract (Python FastAPI and Node TypeScript reference servers ship in the box). Every LLM output is clamped from above by a provable pessimistic ceiling (LpJoinBound — 40.95× tighter than the 2008 AGM bound) so the LLM can never make your plan worse than the engine's native estimate. Transport-floor latency measured at P95 0.07–0.11 ms; live-LLM end-to-end cells honestly marked PROJECTED pending API budget.
Read more
The Kafka + Flink + ClickHouse/Pinot/Druid stack we built between 2018 and 2024 is fragmenting into three forks: single-engine streaming SQL, table-format-as-stream, and OLAP databases that eat the streaming layer entirely. Kafka isn't dying — it's becoming plumbing.
Three months, two abandoned designs, one breakthrough. The one-paragraph version: Apple Silicon GPUs don't have 64-bit atomic_fetch_add until very recent OS versions, and that single missing instruction shapes every other architectural decision in a Metal SQL aggregate engine.
Every analytical engine treats SELECT SUM(x), MIN(x), MAX(x), COUNT(x) FROM t as four passes of the column. Fuse them into a single kernel and the speedup ratio against four-pass code becomes 9x to 25x. Here's why the technique works, and the data shape where it doesn't.
M3 Ultra ships 512 GB of memory at 819 GB/s, addressable by the GPU with zero PCIe transfer cost. Every GPU database project from the past decade was architected around the assumption that memory bandwidth came at PCIe-tax prices. That assumption is now wrong on a fifth of the developer laptops in the world.
samkhya v1.0 ships an LLM-pluggable corrector backend for embedded analytical engines — DataFusion, DuckDB, Polars, Postgres, Iceberg, gpudb. Plug Claude, GPT-4o-mini, or local Ollama into the cardinality-estimation slot via a simple HTTP wire contract (Python FastAPI and Node TypeScript reference servers ship in the box). Every LLM output is clamped from above by a provable pessimistic ceiling (LpJoinBound — 40.95× tighter than the 2008 AGM bound) so the LLM can never make your plan worse than the engine's native estimate. Transport-floor latency measured at P95 0.07–0.11 ms; live-LLM end-to-end cells honestly marked PROJECTED pending API budget.
Every standalone GPU database built between 2013 and 2024 was acqui-hired or pivoted. So why ship gpudb in 2026? Because nobody had wired Apple Silicon's unified memory into a SQL engine — and DuckDB hands you a hundred-thousand-user distribution channel without writing a database from scratch.
Exploring the frontiers of AI, data, and technology. New research and insights published regularly.
About the Author