Research for humanity

MenteE is an umbrella for research aimed at humanity. We work in fields that are not settled yet, and publish what we find.

Scroll

Published research · Open artefacts · Reproducible methods

PP3 preprints
DI2 DOIs
A2Apache 2.0
OWOpen weights
FS41M params, from scratch
GPUSingle consumer GPU
TRIالعربية · English · اردو
FPFailures published
PP3 preprints
DI2 DOIs
A2Apache 2.0
OWOpen weights
FS41M params, from scratch
GPUSingle consumer GPU
TRIالعربية · English · اردو
FPFailures published

The method

Measure first. Then publish everything, including what failed.

MenteE is an umbrella for research aimed at humanity. AI is simply where the answers can be checked today. We take on fields where the answer is not yet settled, build the smallest thing that can be measured, and report the result with the method attached: the numbers, the code, and the recipe that produced them.

Software engineering, where the tooling and the standards for competent performance are still taking shape.

Learning. Living systems. Matter. Medicine. The same approach, applied to questions where more is at stake and verification takes longer.

The order is deliberate. It reflects the scale at which a small lab can work rigorously and verify its own results.

Working this way means publishing collapsed runs alongside successful ones. That is the cost of the method, and it is also the reason the results can be trusted: every MenteE project ships with its method, its numbers, and its complete record.

New preprintZenodo · DOI 10.5281/zenodo.24353178 · Nov 2026

Think while you speak: a model that learns when to think.

AI ships with a binary switch: reason on every input, or none. Our new preprint trains the switch instead. A fast linear path answers by default, a small episodic memory scores the input for novelty, and a deliberate pathway engages only when the answer needs a second look. Pure NumPy, CPU only, gradients verified by hand.

2.8% → 99.2%
Routing gradient

Deliberation rises from 2.8% of easy inputs to 99.7% of medium and 99.2% of hard, recovering the optimal policy at 98.7% routing accuracy.

93.5% @ 67%
Accuracy per compute

Matches an always-deliberate model within 0.2 points while invoking the expensive pathway on only two-thirds of inputs.

94.3% → 73.7%
Memory ablation

Remove the episodic novelty feature and accuracy collapses: without memory, the router sends hard inputs to the path that cannot fit them.

Also newMenteE Code 3B· a coding model built from 53,567 execution-verified agent trajectories, trained for $0.13. HumanEval regression published on purpose, root cause and v2 fix included.Read the report →

Research · mentee-embed

41M parameters, one consumer GPU, and a pretrained baseline beaten on trilingual retrieval.

The uncomfortable question behind mentee-embed is how much of modern NLP's advantage is inherited weights rather than earned capability. So we trained from random initialization — no BERT, no XLM-R, no pretrained backbone — for Arabic, English and Urdu retrieval. v4 reaches 0.874 MIRACL-AR and +146% over v3 on our benchmark, at ~41M parameters and 384 dimensions, fully reproducible.

Two-stage, from zero

Masked language modeling first, then contrastive fine-tuning — a cheap bootstrap that prevents representation collapse.

Honest by default

Measured numbers instead of marketing. Failures documented alongside wins, every artifact on GitHub and Hugging Face.

Low-resource first

Urdu and Arabic treated as first-class languages — not afterthoughts bolted onto an English model.

Preprintmentee-embed-v3· Zenodo · 26 Aug 2026 · CC BY 4.0DOIDownload PDF

Instruments that travel between fields.

The method is not AI-specific. What compounds across fields is the machinery underneath — reading unstructured information, moving it at scale, and surfacing the right decision at the right moment. This is what we rebuild differently in every domain, and the reason the harder questions ahead are reachable at all.

Intelligent automation

Platforms that handle complex, multi-step workflows end-to-end — reducing manual effort and eliminating bottlenecks.

Data at scale

Systems built to ingest, process, and surface insights from massive, heterogeneous datasets in real time.

Decision engines

AI that surfaces the right information at the right moment — powering faster, better-informed decisions.

Document processing

Read, classify, extract, and route unstructured information automatically — at a volume no team could match manually.

Seamless integration

Platforms designed to connect with the tools and systems organizations already rely on — no rip-and-replace.

Built to last

Reliability designed in from the first commit — observability, sane error handling and graceful failure instead of silent breakage.

Under the MenteE umbrella

Chapters of one research programme.

Language, software engineering, work and opportunity — each product is a field the research has reached so far, built and verified by the same two people.

01

Chapter 01 · Language

MenteE Embed

Compact multilingual text embedding model — ~41M params, trained from scratch for Arabic, English, and Urdu retrieval, so search works in the languages a billion people actually use. 384-dim, Apache 2.0.

Read the research →
~41M
Parameters
384
Dimensions
3
Languages
v4
Latest
02

Chapter 02 · Software engineering

MenteE SWE

Beta

Autonomous SWE agent in your terminal. Investigates your repo, plans the change, edits, verifies with your tests, reports with evidence. Model-agnostic, safe by design.

Learn more →
CLI
Terminal-native
Any
Model provider
npx
Zero-install try
MIT
Licensed
03

Chapter 03 · Work & opportunity

RecruAI

Live

AI hiring and interview platform — semantic role matching, ATS-scored CV optimization, unlimited mock interviews with instant feedback, and screening agents for organization pipelines. Built for Pakistan and expanding across the Gulf: UAE, Saudi Arabia, Qatar, Kuwait, Bahrain, Oman.

Explore platform →
AI
Matching
ATS
Scoring
Pipeline
Automation
Gulf
Coverage

The team

Everyone here built what you are using.

There are two of us, and that is still the argument rather than the apology. A small lab that trains from scratch, registers its DOIs and publishes its collapsed runs is a different kind of credible from a forty-person shop that does none of the three. The people answering your email wrote the code and trained the models.

01

Syed Syab Ahmad Shah

Founder & Lead Engineer

syab@menteeai.org

Turns research into systems that survive production.

Owns product direction and system architecture at MenteE. Trained the mentee-embed series from random initialization on a single GPU, published the research openly, and turns it into the agents and products we ship.

5

Publications

41M

Params, from scratch

Zenodo

Open research

02

Sania Shakeel

Co-Founder & Data Scientist

sania@menteeai.org

Makes the data mean something before the model sees it.

Leads data science and machine learning at MenteE — feature pipelines, model evaluation, and the retrieval work that makes our embeddings useful in practice rather than only on a benchmark.

3

Languages

UMich

Applied DS

13

Certifications

We are a small team on purpose. We hire slowly, and we hire people who want to own problems end to end.

Open roles

Blog

Latest articles

Product updates, research notes and founder insights.

Frequently asked questions

The essentials, without the jargon.

Research for humanity, published as we go.

MenteE is an umbrella for research aimed at humanity — and the preprints, weights and code are already public, including the runs that collapsed. If you work in one of the fields we are heading toward, we would rather hear from you than about you.