ALGOS EXPEDITUS
Hi, I’m Satwik.
I build and
experiment with
machine learning.
I’m interested in systems that learn efficiently,
adapt continuously, and make useful sense of messy data.
That has taken me from weightless neural networks and
personalized models to speech representations, GPU computing,
statistical systems,
biological vision, and questions about how intelligent systems
represent, remember, and adapt.
SELECTED WORK
Things that started with a question.
AUDIO + MACHINE LEARNING
Listening beyond the words
What can a voice tell us before we pay attention to the words?
Pitch, rhythm, pauses, and changes in intensity all contribute
to how we hear emotion in speech. I explored how a model might
pick up those patterns through spectrograms, MFCCs, and pretrained
WavLM representations, using acted recordings from RAVDESS.
That brought up another question: would those patterns still
hold when the speaker changed? I kept voices separate across
training and evaluation, then built an app to try the models
with uploaded recordings and a microphone — a way to explore
what they had learned, and where their predictions fell short.
WEIGHTLESS NEURAL NETWORKS
Learning through memory and association
What if a neural network learned less like an optimizer
and more like a memory system?
Weightless neural networks replace continuously adjusted
connection weights with memory-based lookup structures.
Instead of gradually tuning parameters through backpropagation,
they learn by recording patterns and associating new inputs
with previously observed states.
That makes them interesting to me not simply as another neural
architecture, but as a different way of thinking about learning
itself: one centered on memory, association, and state.
COMPUTE
Moving Weightless Networks to the GPU
What happens when an unusual memory-based architecture
meets massively parallel hardware?
I rebuilt and benchmarked WNN workloads using GPU execution,
comparing training and inference performance while monitoring
utilization, memory behavior, CPU load, and thermal telemetry.
DATA SYSTEMS
Making Global Supply Chains Visible
How do you turn enormous economic datasets into something
a person can actually explore?
I worked with environmental input-output data to create
interactive emissions visualizations, Sankey flows,
geographic views, and supporting data pipelines.
RESEARCH
Questions I want to keep pushing.
CONTINUAL + PERSONALIZED LEARNING
What if a model didn’t have to constantly start over?
Many machine-learning systems learn in large centralized
training cycles. New information is collected, aggregated,
and eventually folded back into the model through another
round of training.
I’m interested in systems that can instead learn incrementally
and locally: models that adapt as new information arrives without
repeatedly rebuilding themselves from scratch.
That question led us to explore per-user weightless neural
networks for recommendation — small personal models that
could update continuously as an individual’s preferences changed.
Weightless Neural Networks for Continuously Trainable
Personalized Recommendation Systems
Rafayel Latif · Satwik Behera · Ali Al-Ebrahim
BIOLOGICAL VISION
How does light become something a brain can see?
A photon reaching the eye begins a chain of transformations.
Photoreceptors convert light into electrochemical signals,
retinal circuitry transforms and compresses that information,
and neural pathways carry those signals toward the visual cortex.
I’m interested in understanding that encoding process more
deeply and in how computational systems might eventually
communicate with biological neural systems.
Could an artificial visual encoder produce representations
that biological visual systems meaningfully interpret?
WRITING
Notes on things I wanted to understand.
Machine learning, statistics, computing, science,
and the occasional subject that has absolutely nothing
to do with any of them.
Algos Expeditus started as a place for me to write down
ideas, explanations, experiments, and questions.
The portfolio grew around it. The writing is still the
reason the name exists.
READING
Books currently occupying my head.
Somewhere between bookmarks.
The Myth of Sisyphus
Camus asks how to live in a universe without inherent meaning, arguing for lucid revolt rather than despair.
Psychology in the Indian Tradition
A model of the person built from Indian traditions of mind, consciousness, self, and self-realization.
What If? 2
Rigorous science applied to absurd hypothetical questions, with the answers pushed until they become delightfully impractical.
Up next. Probably.
Complete Poetical Works of Lucian Blaga
A complete English collection of Blaga’s poetry, moving through metaphysics, myth, landscape, and the limits of knowledge.
Pale Fire
A 999-line poem surrounded by a commentator whose notes slowly become a novel of obsession, exile, and unreliable narration.
The Picture of Dorian Gray
A Gothic novel about beauty, corruption, and a portrait that bears the moral cost its subject escapes.
Recently finished.
Atmospheres
Zumthor’s meditation on how materials, light, proportion, sound, and presence create the felt quality of architecture.
Kitchen Confidential
Bourdain’s irreverent memoir of professional kitchens, restaurant culture, ambition, excess, and survival behind the line.
There Is No Antimemetics Division
Science-fiction horror about entities and ideas that erase their own traces from memory — and a war no one can reliably remember.
Piranesi
A man explores an infinite house of halls, statues, and tides while clues accumulate that his world is not what he believes.
A Short Stay in Hell
A condemned man must search an effectively infinite library for the one book containing the story of his own life.
ELSEWHERE
Other places I leave traces.
Hugging Face ↗
Published machine-learning models, starting with my WavLM speech emotion classifier, with weights, evaluation results, and usage notes.
Goodreads ↗
The longer reading list: finished books, current reads, and an ever-growing queue.
ORCID ↗
My persistent researcher identifier and a home for the academic record attached to it.
Google Scholar ↗
Publications, citations, and the research-facing side of my work.
Bay Area · mostly Python · Aries (predictive value: questionable)
