The Work Speaks

Prabakaran Chandran · Columbia University

March 28, 2026 to December 24, 2026 · 270 Days

The Promise

No frameworks. No manifestos. No performance.

270 days of showing up, doing the work, and putting it where people can see it. On Christmas Eve 2026, I launch researchengineer.ing - not as an announcement, but as evidence. Nine months of compounding, visible to anyone who cares to look.

The future state is simple: join one of 5 to 10 frontier AI labs as a Member of Technical Staff or Research Engineer. Not by pitching. By arriving with proof.

How I Work

Like a lab - with rigor, hypotheses, failed experiments, and honest logs.

Like an engineer - if it doesn't run, it doesn't count.

Like a philosopher - because connecting ideas across boundaries is the whole point.

The 20 Output Types

Everything I produce falls into one of twenty types, organized by how resistant they are to self-deception.

Tier 1 - Externally Judged (Can't Be Gamed)

1. Workshop and Conference Papers (Accepted) - The hardest currency. A reviewer said yes. No substitute.
2. Kaggle Profile - Leaderboard positions, medals, competition logs. The number doesn't lie.
3. HuggingFace Profile - Models and datasets with real downloads and usage from strangers.
4. GitHub Profile - Repos that others star, fork, and actually use. Not just green squares.

Tier 2 - Exists Publicly, Quality Is Visible

5. ArXiv Preprints - Ideas staked publicly before acceptance. Timestamped commitment to a direction.
6. Reusable Code Snippets and FOSS - Open-source tools, libraries, utilities that solve real problems for real people.
7. YouTube Presentations - Walkthroughs, paper explanations, technical deep-dives on camera. No hiding behind text.
8. Slides and Courseware - If I built a course on this topic, what would the materials look like? Teaching forces clarity.
9. Frontier ML Systems Projects - End-to-end systems inspired by Stanford-level frontier systems work. Built, benchmarked, documented.
10. Blogs and Research Articles - Writing on frontier ML with genuine, unique perspectives. Not summaries - original thinking.
11. Tutorials and Surveys - Synthesizing a subfield into something a newcomer can actually learn from.

Tier 3 - Real Work, Self-Assessed

12. NumPy, JAX, PyTorch Code Blocks and Tricks - Implementation exercises from scratch. The muscle memory of research engineering.
13. Inference Engineering and Distributed Training Exercises - Serving, quantization, FSDP, pipeline parallelism. The systems side, practiced.
14. Algorithms and Data Structures for ML/DS - 2 to 3 hour focused workouts. Graph algorithms for clustering, flow problems, ML-flavored LeetCode. Design and structure.
15. Hands-on Math Workouts - Derivations in ML, DL, AI methods. Pen and paper, then verify. Optimization, probability, linear algebra, calculus.
16. Programming of Math Concepts - AI creates the scaffolding, I fill in the math. Turning equations into running code, guided by theory.
17. Books Learned and Consumable Content Created - Not just reading - synthesizing books into artifacts others can use.
18. Mental Models on AI/ML/DL Concepts - Distilled conceptual frameworks that compress understanding into transferable intuitions.

Tier 4 - Process and Network

19. Continuous Progress Updates - The daily posting ritual itself. The public log of what happened, not what was planned.
20. Network, Community, Meetups, Hackathons, Collaborations - Showing up in person. Presenting to peers. Building with others. Group projects that compound beyond solo work.

What I Work On

Four channels running in parallel, feeding each other:

Projects - Workshop-level research efforts. World models, causal representation learning, continual learning, interpretability. Real papers, real experiments, real results.

Kaggle and Profiles - Active competition sprints. ADL, Bird Audio, whatever sharpens the blade. Building HuggingFace, Kaggle, and GitHub profiles that show range and depth.

Open Research - Interpretability and RL explorations. Reading, re-deriving, implementing, extending. The frontier, engaged directly.

ML Bricks - The foundation. PyTorch internals, JAX from scratch, distributed training, inference engineering. Math of ML, DL, probability, optimization, statistics, calculus. The things that don't expire.

Applied across: complex dynamical systems, protein and structural biology, LLMs and multimodal intelligence, reasoning, continual learning, and AI for scientific discovery.

The Daily Ritual

Every day. No exceptions.

Posted daily at 9:00 PM EST on LinkedIn, X, and daily.pracha.me.

No fluff. No commentary. No inspirational nonsense. Just the work and the link.

The 3-Day Review

Every 3 days, one question: did the portfolio get heavier?

Measured against the tiers. Tier 1 movement is rare and significant - a paper accepted, a Kaggle medal, a model with traction. Tier 2 should accumulate steadily - preprints, repos, blog posts, videos. Tier 3 is the daily drumbeat - derivations done, code written, exercises completed. Tier 4 keeps the work connected to people and reality.

The portfolio compounds when all four tiers are active. If only Tier 3 is moving, I'm practicing but not shipping. If only Tier 2 is moving, I'm shipping but not deepening. The 3-day check keeps the balance honest.

What This Is Not

This is not a content strategy. This is not a personal brand exercise. This is not a plan.

This is a public commitment to compound quietly for 270 days, across every dimension that matters for the kind of researcher and engineer I want to become. The scoreboard is the work itself.

If you're reading this on December 24, you can see for yourself whether I kept the promise.

---

Day 1: March 28, 2026

Settled · 28 Mar 2026