Zain Ali

Research engineer · VP Engineering & Product, Alexein AI

Imitation can only match its teacher. Experience can surpass it.

  • Six systems built from scratch
  • Three papers one for AAAI-27, one preprint, one in submission
  • 33 results that said no, published

About

Zain Ali standing in front of a chalkboard

I’m Zain. I lead engineering and product at Alexein AI, where we build agents on reinforcement learning and world models, with a memory architecture modelled on human cognition.

I read the paper, rebuild the system from scratch, and report what the numbers said, especially when they said no.

  1. 2026 – nowAlexein AI VP Engineering & Product
  2. 2026ImagineArt Machine Learning Engineer45 → 13 min
  3. 2025 – 2026Victreat Health Tech RL / Agentic Systems Engineer67 % → 91.3 %
  4. 2025 – 2026Machvis Lab, NUST Research Assistant2.6M images · 29K identities

NUST, BS Computer Science, 2022 – 2026. 4th nationwide in the NUST Entry Test. 2nd in ECAT, the engineering entrance test. Top 15 from Pakistan at the International Mathematical Olympiad.

Selected work

  1. gedankenEinstein’s 1907 step, rebuilt as a test for AITimes it checked its own leap, unprompted2/140
  2. whetstoneSelf-improvement, audited for reward hackingAccepted gains that only gamed a visible test39.6%
  3. lociA zero-fit theory of where a memory is filedMean error predicting 156 runs it never saw0.009

Also built, and written up in full: engram (memory, modelled on the hippocampus), spindle (an inference engine from scratch), reverie (planning inside a learned world).

33 things that didn’t work, each with the number that showed it

Papers

  1. Imagining without CheckingDoes a language-model agent check an equivalence before extending it?Prepared for an AAAI-27 workshop, February 20270.993
  2. CONOIDAgent teams picked by track record, not by self-description.Preprint, 2026 · with Muhammad Ahmed Mohsin66%
  3. Sheaf ReputationThree reductions of one cellular-sheaf constructionIn submission
Read the CONOID preprint

AI has learned to predict the next word. I work on the harder problem: choosing the next action.

Interests

What I work on, and the position I hold on each.

  1. Reinforcement learning

    A reward is a specification, and its loopholes are the first thing a strong policy finds. So I write the check before I train the policy.

  2. World models

    A plan is only as good as the worst prediction along it. Find where the model goes blind before trusting anything it plans.

  3. Combinatorial search

    In an exponential space the heuristic is the algorithm. Learn the heuristic; keep the search sound.

  4. Agentic AI

    An agent is a loop, and a loop compounds whatever its verifier misses. The verifier, not the model, sets the ceiling.

  5. Memory

    A context window is a buffer, not a memory. Memory is three decisions, what to keep, where to file it and what to forget, and each one can be measured.

  6. LLM engineering

    Single-stream decoding is bound by memory bandwidth, not FLOPs. The speed lives in the KV cache and the scheduler.

Bring a hard problem and a way to check the answer.

zali.bscs22seecs@seecs.edu.pk
GitHub LinkedIn