Hi, I’m Akram

I’m Mohammed Akram, a computer science undergraduate at B.S. Abdur Rahman Crescent Institute of Science and Technology (GPA 9.77/10, rank 1 in batch). I study how language-model systems can improve themselves without changing their weights, and what that improvement costs.

Right now I’m working on self-harnessing recursive language models at the Algoverse AI Research Program with Dr. Simon Coumes: can a frozen model mine its own failures, rewrite the harness that controls its recursion, and still generalize to inputs 8–32× longer? Separately, I measured the energy and carbon cost of fine-tuning a reasoning model: fine-tuning a strong instruction-tuned model on GSM8K spent 2.99 kWh across 10 runs and lowered accuracy by 23–28 points.

Before research, I built AI products. I co-founded Airbil, an AI-native agency building multi-agent systems, and PocketLink, a link-in-bio platform that grew to 20,000+ users. Building agents in production is what got me interested in why they fail.

Interests: recursive self-improvement · agentic systems & harness engineering · recursive language models · long-horizon reasoning · efficient AI

I’m applying to graduate programs and am open to research collaborations. Reach me at [email protected].

Research

In preparation · experiments ongoing

Self-Harnessing Recursive Language Models

Manuscript in preparation; targeting an AAAI 2027 workshop

Can a frozen recursive language model mine its own failures, rewrite its own harness, and keep working on inputs 8–32× longer than it was optimized on?

Presented · completed

The Carbon Cost of Reasoning

National AI Summit on Industry 5.0, 2026 (Paper ID AIS 052)

Measuring the energy and CO₂ cost of LoRA and QLoRA fine-tuning on Gemma 4 for math reasoning. Every run made the model worse: 10 runs, 2.99 kWh, and a 23–28 point accuracy drop.

All research →

Publications & presentations

Education

Experience

Full details in my CV.

Recent writing