Finishing the Last 1%
Preamble
There is a common rule of thumb in software and research known as the Pareto Principle: the idea that 80% of a project’s results come from 20% of the effort. What I think is less obvious about the remaining 20% (and particularly that final 1%) is that it demands an asymmetric amount of energy, discipline, and attention to detail. In the phase where an idea transitions from an internal prototype into something capable of standing up to external scrutiny, every decision you made flippantly in the first 80% of your project comes back to haunt you and analysis paralysis is a real risk to finishing up strong.
This month, that reality hit me all at once. I found myself navigating the closing stretch of three major milestones simultaneously: finishing my spring internship at Brookhaven National Laboratory, rushing to polish a year-long research project to for my first AI conference submission, and preparing to walk across the stage for my undergraduate graduation. Each required a distinct final sprint, but looking back at them reveals how differently that last 1% can manifest depending on what you are trying to complete. In this post, I wanted to catalog some reflections on these final pushes.
Part 1/3: BNL ~ Granularity for the Final 5%
My time at BNL working in the climate department was a noticeable change of pace from standard academic research and my previous experiences in machine learning. It was a structured, consistent 40-hour work week where I helped co-manage a project divided into three distinct phases. The core task was spatiotemporal time series inference: analyzing 3D physical atmosphere data as it evolves over time to be able to more accurately detect cold fronts with limited sensors.
Working on climate data taught me that even when dealing with massive datasets, the quality of downstream machine learning tasks relies entirely on understanding the physical phenomenon at the most granular level, down to the exact minute an event occurs. I had to put myself in the shoes of a meteorologist and ended up often looking minute by minute through some time series data which spanned over a year of collected data even after implementing my statical and machine learning methods.
[Stage 1: Identification] ──> [Stage 2: Visualization] ──> [Stage 3: ML Inference]
Sliding Window Statistics - Composite & Correlation Plots - Cold Front DetectionBecause of this, I spent my final 5% of time at BNL semi-obsessively refining the first two stages of our pipeline rather than rushing to tweak the model architectures in the third step. An abridged description of the three stages of the project were:
- Stage 1: Identifying cold front events via sliding window statistical measures.
- Stage 2: Generating composite and correlation plots to visualize the weather dynamics, allowing us to make precise, micro-scale timing adjustments to event boundaries.
- Stage 3: Anomaly detection utilizing softmax regression trained on verified ground truth.
This gave me first-hand experience with “data-ing”, a concept I learned about related to OpenAI’s development of models like GPT-3, where painstaking focus on baseline data design proved more critical than excessive tweaks to the language modeling design itself. I learned that ensuring the inputs and event labels were near perfect in Stages 1 and 2 guaranteed the success of the ML inference in Stage 3. This was a lot of work but a good reminder about discipline starting at the foundation of an endeavor.
Part 2/3: MIT ~ the Marathon and the Sprint
In contrast to the structured schedule at BNL, my research with MIT was a long-scale marathon. I had been working on the project since June 2025 with months of floating ideas, shifting research questions, and testing prototypes before targeting my first AI conference submission. We were focused on understanding LLM internal representations by using program synthesis to generate clean Python code proxies for attention matrices.
After nearly a year of work, the final two weeks before the deadline turned into an intense sprint due to code debt. Months of exploring different approaches had left our codebase fragmented. To complete the project, we had to execute a “Pareto rebuild.” We archived the patchwork of historical scripts and rebuilt our core library from scratch using a single, unified strategy for program synthesis locking down strict prompt templates, refinement loops, and curated examples.
This rebuild became the final gating item. Everything else from running our final experiments to editing the final manuscript and making sure our figures were aesthetic had to wait until the core foundation was clean and unified. The last 1% here was about taking a step back to rebuild, ensuring the empirical results were built on solid ground with a clear narrative.
Part 3/3: Graduation — An Inflection Point
Then came graduation. Reaching the end of my undergraduate studies made me reflect on the foundation that brought me here. Before starting at NJIT, I went through an 8-semester sprint in only two years to complete three Associate Degrees in Physics, Computer Science, and Mathematics. That was a relentless, max-credit push through summers and winters that left me pretty tired at the end. Then at NJIT, pursuing the BS/MS and Honors program simultaneously was slightly less intense but still a sprint at times.
My family often tells me I should pause and celebrate accomplishments more as I’m not very big on celebrating these milestones. But I have always appreciated this trait in myself because I think a mindset that prevents you from dwelling on success is the same mindset that prevents you from dwelling on failure. Regardless of the external circumstances, you can just appreciate the work and the lessons and move on to the next challenges where you’ll be a little bit better than before.
As the final weeks turned to days, and the 5% left became 1%, the nature of the sprint shifted. Instead of pulling all-nighters or grinding through code, I spent my final days before graduating resting, putting together a new fixed gear bike from parts I bought at the Brooklyn Bike Jumble, and just relaxing.
I’d like to think this last month marked an inflection point in my academic path where I can transition from a improving at learning and absorbing, to improving my ability to create. My sprints four years ago were about proving I could pass tests and do well on coursework. My sprints today are about building original tools, proving research hypotheses, and trying to create things that did not exist before (and that people might actually care about).
While these three conclusions looked slightly different on the surface, writing this outlines their pretty similar theme. Bring a resolution to an endeavor in all cases required having a strong baseline and revisiting those foundations towards the end, in order to make sure that conclusion was built from the same great ideas that initated the journey in the first place. This is something I’ll take with me this summer to IBM Research and as I wrap up my Master’s degree next semester.
Want to be notified for the next blog?