august updates, 8/2/26, gainesville:
hey, i’m jatin, and i build ai systems to understand what actually makes them fast.
i’m studying computer science at the university of florida. most days, i’m somewhere between implementing decoding algorithms, staring at cuda timelines, and asking whether fewer model calls really translate into lower wall-clock latency.
my work sits at the intersection of llm inference, gpu performance, and reliable evaluation. i care about exactness first: correct sampling, reproducible runs, clean evidence, and measurements that survive beyond a good-looking benchmark.
things i like: profilers, careful ablations, long walks, and systems that explain themselves.
things i don’t like: benchmark claims without bottleneck analysis.
i’m jatin, an ai systems performance engineer based in gainesville, florida.
i’m currently building reproducible clinical llm evaluation infrastructure at the university of florida college of medicine using slurm, apptainer, python, neo4j, and graphiti.
i’m also studying latency-aware speculative decoding: a correctness-first engine with kv caching, rejection correction, draft-model distillation, and gpu-level profiling on nvidia l4s.
previous stuff i did
- evaluated agent memory against 1,682 factsbuilt a six-stage verification pipeline for temporal clinical knowledge-graph memory.
- published an acl 2024 long paperbuilt and evaluated multimodal fusion for clinical document summarization.
- served 1m+ daily ml inferencesbuilt production systems at icici bank and improved p95 serving latency by 5%.
- increased click-through rate by 9%developed feature-store ranking and propensity workflows for banking products.
- cut target-model passes from 64 to 22implemented a correctness-first speculative decoding engine reaching 2.91 tokens per target pass.
- shipped into joomla 4 corebuilt the guided tour feature for a cms used by more than 2m sites, then returned as a mentor.
- won two national-level buildscolumbia × aws bedrock hackathon winner and smart india hackathon winner.