# Applied ML and research

> I understand the models underneath, not just the APIs on top.

An MSc in AI with reinforcement learning research, five peer-reviewed papers, computer vision, and fine-tuning small models with QLoRA. It is why I can reason about failure modes, evaluation design and when a model is the wrong tool.

- Core skills: Python
- Domains: Research and public sector
- URL: https://farhanbaluch.com/capabilities/applied-ml-research

## Case studies

- [Graph-augmented reinforcement learning for stock trading](https://farhanbaluch.com/work/graph-rl-trading.md) (Sep 2022 – Sep 2024, Graduate research, quantitative finance): Researcher, MSc thesis. An RL agent is a sequential decision-maker, like an LLM agent. Designing rewards, comparing a policy against baselines and explaining what it learned is the same discipline I now bring to agent evaluation.
- [Applied AI prototypes, from crops to small models](https://farhanbaluch.com/work/applied-ai-prototypes.md) (2023 – present, Public sector, small businesses and independent work): 25% → 75%, exact-match extraction accuracy after QLoRA fine-tuning of Llama 3.2. The models underneath agents: vision, retrieval and fine-tuning. Knowing when a small fine-tuned model beats a large prompted one, and how to measure it, is what lets me choose the right tool on client work.

## Evidence from the log

- 2025-07-07: **Grounded estimates in historical projects.** Added retrieval over past estimates with Azure AI Search and a regression baseline on Databricks, so every number comes with a confidence interval and citations. AI angle: Grounding keeps the LLM from inventing numbers. It adjusts a statistical baseline and cites the past projects it used. [Knowledge, retrieval and context; Applied ML and research] (https://farhanbaluch.com/work/estimation-platform.md)
- 2024-09-03: **Completed an MSc thesis on graph-augmented reinforcement learning.** A TD3 trading agent with Node2Vec embeddings of how stocks relate, which outperformed RL baselines on Dow 30 data, plus visualizations that explain its strategy. AI angle: An RL agent is a sequential decision-maker, like an LLM agent. Designing rewards, comparing against baselines and explaining behavior is the same discipline I now apply to agent evaluation. [Applied ML and research] (https://farhanbaluch.com/work/graph-rl-trading.md)
- 2024-01-15: **Took small-business AI use cases from problem to prototype.** At the Vector Institute, helped small and mid-sized businesses define an AI use case and reach a working prototype in three months. AI angle: Early practice in what solution work really is. Most of it is scoping the problem so a model can actually help. [Solution architecture and delivery; Applied ML and research] (https://farhanbaluch.com/work/applied-ai-prototypes.md)
- 2023-05-15: **Built crop-disease detection and a policy chatbot.** At Agriculture and Agri-Food Canada, built computer vision for early crop-disease detection and a chatbot that answers farmers' policy questions. AI angle: Vision models and a retrieval chatbot over policy documents. My first RAG system, in a public-sector setting. [Applied ML and research; Knowledge, retrieval and context] (https://farhanbaluch.com/work/applied-ai-prototypes.md)
- 2022-09-06: **Started graduate research on reinforcement learning for trading.** Joined the University of Windsor as a graduate research assistant, working on RL agents for portfolio trading. AI angle: The research foundation in agents, reward design and evaluation that the rest of this log builds on. [Applied ML and research] (https://farhanbaluch.com/work/graph-rl-trading.md)