Reinforcement learning for routing
1.6% from optimal on TSP20
Optimality gap vs. LKH across 1,000 test graphs. Lower is better.
This work Classical baselines
Open to roles · Summer & Fall 2027
Shazan Ansar Mohammed
Data Scientist AI Architect Data Engineer
Data scientist specializing in NLP and large language models, with a data engineer's discipline for pipelines and an architect's eye for systems. Published research on LLM-based machine translation at RANLP 2025 and ICAI 2026.
Proven results
Measured, not claimed.
Results from published research and reproduced experiments, and the pipelines behind them.
LLM post-editing · ICAI 2026
Fine-tuning quality, no retraining
Share of the fine-tuning gain recovered, across four domains.
- Domains
- 4
- Retraining
- None
- Recovered
- 87–92%
Pipelines
How each result was produced
English–Nepali clinical translation
PublishedRANLP 2025- 1
Clinical textPython
English–Nepali medical domain
- 2
70k+ sentence pairsPython
Cleaned and aligned corpus
- 3
Fine-tune two modelsPyTorch · HF
mBART and NLLB-200
- 4
6 metricsMetric suite
BLEU, chrF++, METEOR, BERTScore, COMET, perplexity
- 5
Errors and ethicsHuman review
Error profiling for clinical use
70k+
sentence pairs
2
models fine-tuned
6
evaluation metrics
Outcome. NLLB-200 was more accurate and hallucinated less on clinical text.
Read the case studyProjects
Flagship work.
- AI Architecture
LLM Post-Editing for Domain Adaptation
A frozen translation model drafts; an LLM adapts the draft to the target domain at inference time.
87–92% of fine-tuning quality, zero retraining
- LLMs
- NMT
- Evaluation
- Data Science
Clinical Machine Translation
Fine-tuned mBART and NLLB-200 for English–Nepali medical text, evaluated on six metrics with an ethics review.
70k+ sentence corpus · RANLP 2025
- PyTorch
- Hugging Face
- COMET
- Data Science
Reinforcement Learning for Routing
Reproduced Google Brain's pointer-network approach to the travelling salesman problem on an A100.
1.6% from optimal; critic baseline halved the gap
- PyTorch
- RL
- A100
- Data Engineering
Real-Time Capital-Markets Platform
Live market data streamed through Kafka into a medallion lakehouse that produces positions and PnL.
Bronze → Silver → Gold, streaming to analytics
- Kafka
- DuckDB
- FastAPI
- Data Science
Transformer Architecture Study
Controlled ablation of feed-forward designs in a GPT model: width, activation, and gating.
Width beat activation; SwiGLU most parameter-efficient
- PyTorch
- GPT
- Ablations
- AI Architecture
RAG Analytics Assistant
Grounded question-answering over business data with filtered retrieval, routing, and caching.
Grounded answers with controlled latency and cost
- RAG
- Embeddings
- Caching
Research
Published research.
Large language models for machine translation, evaluated with care.
Accepted, to appear2026Large Language Model-Based Automatic Post-Editing for Test-Time Domain Adaptation in Neural Machine Translation
ICAI 2026
Read the summary
Under review2026Large Language Models for Neural Machine Translation: An Evidence-Graded Survey of Methods, Evaluation, and Deployment
Journal of Universal Computer Science (J.UCS)
Published2025Advancing Clinical Translation in Nepali through Fine-Tuned Multilingual Models
RANLP 2025
Read the summary
Thesis · in progress
TRACE-MT
When does retrieval truly help LLM translation, and when is it just priming?
Grant project · in progress
LLM post-editing for Tagalog and Spanish
Norris Research Program, Fall 2026.
People
Built with great people.
Advisors, co-authors, and the researchers I mentor.
- BADr. Benyamin AhmadniaResearch advisor · Thesis chair · CSUDH
- SHDr. Sahar HooshmandCo-author · Thesis committee · CSUDH
- MRDr. Marsa RayaniThesis committee · CSUDH
- LZDr. Li ZhouCo-author · Harvard Medical School
- SSSumaiya ShaikhCo-author · CSUDH
- BPBibek PoudelCo-author · CSUDH
Skills
Built for three roles.
55 skills, organized by role, from advanced to foundational.
Rigorous modeling, experimentation, and evaluation.
Advanced
- Causal & counterfactual inference
- Hierarchical & mixed-effects models
- Experiment design & power analysis
- LLM fine-tuning & evaluation
- Deep learning (PyTorch, TensorFlow)
- Reinforcement learning
- Model interpretability (SHAP)
- Ablation studies
Core
- Supervised & unsupervised ML
- Feature engineering
- Time-series forecasting
- NLP & transformers
- Hypothesis testing & bootstrap CIs
- A/B testing
- scikit-learn, Pandas, NumPy
Foundations
- Python, R, SQL
- Statistics & probability
- Linear algebra
- EDA & visualization
- Tableau, Matplotlib, Seaborn
Experience
Experience & credentials.
Graduate Research Assistant, NLP and machine translation
Jan 2026 – presentCalifornia State University, Dominguez Hills
Research Assistant, clinical NLP
Feb 2025 – Aug 2025California State University, Dominguez Hills
Data Science in AI & ML Intern
Mar 2023 – Nov 2023Code for India Foundation
M.S. Computer Science
2024 – presentCalifornia State University, Dominguez Hills · GPA 4.0
- AWS Certified Solutions Architect – AssociateAmazon Web Services · Aug 2025 – Aug 2028
- Google Data Analytics SpecializationGoogle · Oct 2023
- Machine Learning for Trading SpecializationGoogle Cloud · Oct 2023
- IBMPython for Data ScienceIBM · Jun 2022
LeadershipProgram Manager, Idea Incubator (R&D)
Mahatma Gandhi Institute of Technology, Hyderabad
Life
Off the clock.
- Reading
- Tennis
- Hiking
- Cafés
- Travel
- Films and TV
Drag to spin
Speaks English, Hindi, Urdu, Telugu, Arabic.
Contact
Let's talk.
Hiring a data scientist, building a startup, or working on language AI?