Prabakaran Chandran
Applied research and engineering — AI, machine learning and data science.
பிரபாகரன் சந்திரன் · Tamil Nadu → New York
I take ideas in AI, machine learning and data science from theory to implementation, and on to production systems — through continuous experiments and improvements, on real-world problems.
Nearly eight years of hands-on work, from control engineering through data science and machine learning to research — Mu Sigma, Captain Fresh, Informatica, Nile — carrying one conviction the whole way: anything that can be seen as a system can be modelled, understood and acted upon.
Finishing an MS in Data Science at Columbia, and teaching as a TA in two courses: Foundation Model Engineering, and AI in Practice, with its focus on AI and games.
Open to Applied Scientist and Research Engineer roles from December 2026.
From theory to production
All builds →I frame the problem, prove the method experimentally, and take it from experiment to a system people use. Each piece of work below is marked by how far along that line it got.
- 01TheoryWhy it works, from first principles.
- 02ImplementationCode where every step can be inspected.
- 03ExperimentsAgainst a baseline, one change at a time.
- 04ProductionRunning on a real problem, for real users.
Filled marks show how far each piece of work got. Red is where it stands now; a red ring means prototyped, not yet in use.
Research · Columbia · CRIS Lab
- Learning without forgettingTeach a model cats, then dogs, and it forgets cats. A fix that keeps none of the old images — now prototyped for deployed agents.
- A world model for what-if questionsIt infers hidden masses from a video of a scene, then plays the scene forward from a different start.
- What you want, not what you were shownRecommenders learn from clicks the platform itself caused. Separating that from genuine preference.
- Agents that cooperate with strangersTwo robots trained apart meet in one kitchen. What makes a shared convention obvious to both.
- Networks that obey physicsA simulated pendulum gains energy from nowhere. A network built so that it cannot.
In production · Industry
- Every shrimp pond on India's coastCaptain FreshMapped from satellite images, with a loss derived for the case every model got wrong — dry ponds that look like farmland.
- Fixing failed data pipelinesInformaticaA multi-agent system support engineers self-serve from when cloud data-integration jobs fail.
- Medical case files into reportsHealthcare startupThousand-page medical and insurance files turned into reports and patient chronologies — from founding team to beta.
- Checking AI agents before releaseNileAn evaluation gate that tests agents' answers before users see them; the team scaled it into the product.
- Solar power forecasts for tradersMu SigmaWeather-corrected forecasts with honest uncertainty, used to size intraday bids.
What I work on
Intelligence is becoming abundant. I work on what happens downstream of it — which problems are now worth attempting, and what it takes to build the systems that attempt them. The ones I keep returning to are the hard ones: human–AI coexistence, complex systems, safety and prosperity, and advanced AI turned on disease, pandemics, wellbeing and social problems.
- AI safetyWhere advanced AI systems fail, and how to catch it early. Course →
- Causal inference & statisticsTelling cause from correlation, so decisions rest on what actually works. AI →
- Complex systemsHow interacting parts produce behaviour none of them has alone. Course →
- Learning systemsReinforcement learning, representations, and models that keep learning. AI →
- The long viewHealth, wellbeing, and how people and AI live together as AI keeps improving. Problem Solving →
- Sep 2026Preferred Problem Spaces — Building My Career Around What Keeps Me CuriousMusing
- Sep 2026The Next Starting LineMusing
- Sep 2026Complex Systems, Not Integrations: What Forward-Deployed/AI Engineering Is MissingArticle
- Jul 2026Memories 1: Growing up in SingalandapuramMusing
- Jul 2026Pre-NEET: the story the data missed, and how people misread Tamil NaduMusing
- Jul 2026Curriculum as a Living SystemMusing
- Jul 2026How I Built Myself a Curriculum From Zero (thinking out loud with an AI)Musing
- Jul 2026A State of Me, and What's NextMusing
Questions I am working on
- What does a generalist become in the age of AI, and where does that still matter?
- Which problems are worth focusing on under superintelligence — and how would we know?
- Where do misalignment, malfunction and security threats actually emerge in advanced systems?
- What holds epistemic sanity in place when knowledge gets cheap to produce?
The ground I stand on
Why →The world is complex — interacting parts, feedback, adaptation, emergence. Complexity science is the ground: the superset of dynamics, evolution, entropy and computation, and the interactions between them. Not a lens I picked, but how the world works — and where advances in AI, including AGI, have to be understood from. It is also why I think across disciplines rather than inside one.
- Dynamics & controlHow systems change over time, and how feedback holds them together or tears them apart.
- Information & computationWhat can be known, compressed, learned and computed — and at what limit.
- Entropy & energyWhat order costs, where systems tend, and why some structures persist.
- Evolution & adaptationHow systems vary, select, learn and keep working in a world that moves.