I build intelligent systems that read the world and act on it - from large language models and agentic pipelines to financial ML and production NLP. I care less about collecting tools and more about the reasoning behind a system: why an architecture holds, where it breaks, and what a benchmark actually proves.
| Track | What I'm doing | Depth | |
|---|---|---|---|
| 📄 | Reading | Agent architectures - ReAct, Toolformer, Plan-and-Solve, Reflexion | ███████░░░ |
| ⚡ | Building | Financial ML - stock valuation, reverse-DCF; LLM agents | ████████░░ |
| ✍️ | Writing | Weekly ML research breakdowns - papers, made accessible | █████████░ |
# the problems I keep coming back to
# tools I reach for by default
# the ML/AI libraries I ship with
[2026-02] NLP Specialization - DeepLearning.AI
[2026-02] Deep Learning Specialization - DeepLearning.AI
[2026-02] Machine Learning Specialization - DeepLearning.AI
[2025-10] Mathematics for ML & Data Science - DeepLearning.AI
[2026] AWS Fundamentals - Accenture
[2026] GenWizard Generative AI - Accenture
+ 38 more
| Type | Titles |
|---|---|
| 📕 Books | Deep Learning (Goodfellow) · Designing ML Systems (Huyen) · Hands-On ML (Géron) |
| 📄 Papers | Attention Is All You Need · ReAct · Toolformer · Reflexion · DPO |
| 📰 Blogs | OpenAI · Anthropic · Hugging Face · Lilian Weng · The Gradient |
Now reading Reflexion: Language Agents with Verbal Reinforcement Learning ███████░░░ 70%
Weekly ML research breakdowns - turning papers into something engineers can actually use.
sujay@ml ~ $ posts --latest --platform=linkedin
[AGENT] AI agents in 2026 can't plan - this paper might fix that
[TOOLS] Tool-calling: the hardest unsolved problem in agentic AI
[FINML] ML for finance sounds powerful - here's the gap nobody mentions
[SCALE] 1M-token context windows: what they actually mean for engineers
sujay@ml ~ $ _

