rishi more

rishi more

理志

making language models faster, cheaper, and easier to trust.

base
sunnyvale, california
prev
intersystems, apple ml, jhu dsai, mastek
uni
johns hopkins (cs)
focus
teaching llms to behave, then making them fast

selected work

conformal thinking

making llms lazy in a good way: think hard only when it matters.

reasoning models that stop once they're sure. a dual-threshold early exit ends the thinking when the answer is confident or the reasoning has stalled, which cuts reasoning tokens by 12% at matched accuracy.

read the paper
venue
icml 2026, main track
role
co-first author
with
apple ml research, jhu dsai
tested on
4 reasoning models, 4 benchmarks, 2.7k+ problems

kv roofline

fitting 580k tokens of context on an 8gb gpu.

int4 kv-cache quantization with a fused triton decode-attention kernel. kv memory drops 3.47x, and decode runs up to 27x faster than pytorch at 128k context.

triton, pytorch, cuda

view code

post-training failure evals

finding where "aligned" models secretly get worse.

an evaluation harness that walks a base model through sft and dpo, then pins each regression on the step that caused it. six failure modes, from reward hacking to sycophancy and false confidence.

python, pytorch, llm judge

view code

more research and projects

experience

  1. summer 2026

    intersystems

    sde, ai experience

    rlhf on qwen2.5-32b with clinician preference data, 93% on a clinician-workflow golden set.

  2. 2025–26

    apple ml research + jhu dsai

    grad research, llm inference

    co-first author of conformal thinking, and built the vllm calibration and evaluation pipeline behind it.

  3. 2022–23

    mastek

    mle co-op

    edge voice authentication with sub-100 ms p99 latency and 96.6% accuracy.

off the keyboard

  • boulderingV9
  • chess1700
  • liftingbench 250lbs @ 118lbs
  • baking72-layer croissant

get in touch

leave a note here, or email me directly.

rishi@koi terminal