ML Engineer & Researcher | Operations → AI | Efficient Architectures for Reasoning
I optimize systems under constraints. Whether managing supply chains or training neural networks, the goal is the same: maximum performance, minimum waste.
Background: 10+ years in operations and supply chain → systematic pivot to AI/ML engineering and research (2024–present)
Research focus: I'm interested in how intelligence emerges from structure rather than scale. Specifically: architectures that reason iteratively rather than in a single forward pass, where the answer emerges from an internal simulation, not a lookup. My current thread runs through object-centric representations, state space models, and geometric deep learning as building blocks for more efficient, interpretable reasoning systems.
Grade A (9.4/10) Tamil-language agricultural instruction dataset for smallholder farmers — built from 679MB of Kisan Call Centre government data, TNAU extension guides, and ICAR contingency plans. One of the only agricultural datasets in the world with farmer mental health crisis routing.
Solo builder — data collection, pipeline engineering, metadata design, platform optimization
- 10 iterative submissions across 6 weeks; discovered that metadata specificity (not row count) drives dataset quality
- 187 rows, 20 categories, 48 crops, 17 structured columns, zero nulls
- Received the first honorary award from Sara Hooker (co-founder, Adaption Labs)
- Tech: Cohere · Adaption Adaptive Data · Python · pandas
- Dataset: HuggingFace | Repo: GitHub
Privacy-first, fully offline cross-lingual document QA for migrants and newcomers. Upload a foreign legal document, ask questions in your own language, get answers with source citations and a hallucination trust score — entirely on-device, nothing sent to the cloud.
- Pipeline lead and eval researcher on a 7-person team
- Built the full RAG pipeline: PDF extraction → BGE-M3 embeddings → ChromaDB retrieval → Tiny Aya 3.35B generation → mDeBERTa NLI hallucination check
- 9,063 automated evaluations across Chinese, Hindi, and Polish
- Key finding: the bottleneck in cross-lingual RAG isn't the language model — it's the embedding space mismatch between question and document language
- Tech: Tiny Aya GGUF · llama-server · BGE-M3 · ChromaDB · mDeBERTa · Gradio
- Repo: Docunative | Cohere-Labs-Community/docunative
Visualizing neural network optimization surfaces to understand why some architectures train better than others. Tech: PyTorch · matplotlib · 3D surface plots
Implementing sequence models from scratch: RNN → LSTM → Transformer → Mamba. Analyzing gradient flow and computational complexity at each step.
Goal: understand the evolution of sequential architectures from first principles — not just that they work, but why they work, and what each generation was designed to fix. Tech: PyTorch (no high-level libraries) · comparative benchmarking
ML Engineering Intern | M2M Tech (2024–2025)
- Built and deployed end-to-end ML pipeline (XGBoost) for a client startup
- Feature engineering, model training, API deployment, monitoring
- Stack: Python · FastAPI · Docker · AWS
Supply Chain & Analytics (2017–2023)
- Optimized inventory systems — prevented $500K in stockouts
- Built predictive analytics for demand forecasting (+25% revenue impact)
- Automated reporting pipelines (saved 15 hrs/week)
Deep Learning: PyTorch · Transformers · State Space Models (Mamba) · Computer Vision · NLP
ML Engineering: Docker · FastAPI · CI/CD · AWS · Hugging Face
Foundations: Linear Algebra · Optimization · Probability Theory
M.A.Sc. Electrical Engineering | University of Windsor (2012)
B.E. Electronics & Communication | Anna University (2008)
Recent Training (2024–2025)
- University of Toronto: ML & Data Science Professional Certificate
- Stanford: Machine Learning Specialization (Andrew Ng)
- Google: Advanced Data Analytics
The Meta Gradient — technical writing on architecture evolution, deep learning fundamentals, and the philosophical rabbit holes these lead me down.
"Intelligence emerges from constraints, not just compute."