Current RLVR defines state as unbounded token history — a degenerate MDP formulation that prevents RL from discovering genuinely new reasoning strategies. Motivated by the NeurIPS 2025 oral showing this capability ceiling is structural, not a data or compute problem. Replacing history-as-state with a VAE-based latent belief over reasoning trajectories: a compact, learned Markov state that captures where the agent is in the solution space, not what it said. Latent variance used as an intrinsic exploration bonus. Ablating against standard GRPO and the Markov state paper (arXiv 2603.19987).
github.com/MohammedYaseen97/latent-markov-world-model ↗Latent world models for abstract problem-solving — an agent maintaining a diffusion-based belief over solution-space position, navigated by RL-driven epistemic uncertainty reduction. As distinct from physical-world models (JEPA, LeWM), this targets the epistemic state of a reasoning agent. The VAE project is the tractable proof-of-concept; the diffusion model is the PhD.
Gymnasium GPU scheduling benchmark on a discrete-event simulator with realistic bursty workloads and hard feasibility constraints. PPO end-to-end with action masking. Controlled ablations on two research contributions: cross-attention policy vs MLP, and lookahead representation design (off / per_gpu / sorted / CDF). Rigorous eval hygiene — deterministic eval, best-checkpoint selection, separate VAL vs held-out REPORT seeds.
Multilingual GPT-2 from scratch — BPE tokenizer, pretraining on 40GB Indic text, SFT, and DPO from primitives. Key finding: raw perplexity is an invalid cross-tokenizer metric. Bits-per-byte normalisation reverses the ranking entirely. 4.6× compression improvement on Kannada, 2.75× on Devanagari vs tiktoken. DPO increased chosen log-prob +6.1%, decreased rejected −16.8%.
Standalone exploration of Byte Pair Encoding tokenizer design with Indic-aware Unicode split patterns. Isolated study of how tokenizer choices affect downstream compression and model performance — the findings that fed into the Indic GPT-2 project's core methodological contribution.
MSc dissertation. Empirical study of how Transformer encoders represent idiomatic meaning. Adapter-based transfer experiments probing cross-lingual idiom generalisation (English ↔ Portuguese). MirrorWiC-style fine-tuning pipeline on HuggingFace with BERT/mBERT baselines.
Designed a cognitive model of human attention and decision-making grounded in Kahneman's System 1/2 framework; modelled decision processes as Reconfigurable Bayesian Belief Networks spanning domains, time periods, and psychographic profiles. Architected a research-quality benchmark evaluating synthetic humans against real humans across six behavioural dimensions at individual and population level.
Multilingual GPT-2 from scratch; LLaVA clone (CLIP ViT + Phi-2); deep study of MoE, RoPE, KV caching, speculative decoding, flash attention.
Planner Agent using Tree-of-Thought + LangGraph for multi-step tool routing. Hybrid search (Qdrant + BM25) over Google Drive integrated into RAG pipeline. Experiments across DSPY, Outlines, Instructor for prompt tuning and structured output.
Built and productionised a shopping assistant chatbot reaching #1 on Product Hunt. Reduced TTFT to 3–5s and cut operational costs ~80%. Hallucination detection, knowledge graph retrieval, transformer-based search pipeline.
Speaker diarisation via novel attention-based clustering module. VQ-VAE and attention-CNN benchmarks on DIHARD challenge datasets.
Machine Learning, NLP, Text Processing, Parallel Computing / CUDA
Algorithms, Machine Learning, Artificial Intelligence, Big Data Analytics