Hi, I'm
V SS Karthik
▍
B.Tech CS (AI/ML) student who lives in the “how does this actually work?” phase. I build things from scratch, from write-ahead logs and skip lists to models that explain their own failures.
01 About
I'm drawn to the layers underneath the tools we use every day: how a database survives a crash, how a hash map really behaves under load, why a model fails on the inputs it fails on. I like writing the thing myself first, then making it fast.
- Studying B.Tech CS (AI/ML)
- Most comfortable in Python, then C
- Based in Telangana, India
- Motto “Low-level logic, high-level curiosity.”
02 Projects
Selected work, all open source. Click a card to see the code.
Systems
CDSA
A zero-dependency, memory-safe C data structures library (Vector, HashMap, SkipList, ART, RingBuffer and more). It also powers a CPython extension for Kedis-Python's storage layer.
kedis-python
A Redis-inspired in-memory store with a custom binary wire protocol (KESP), asyncio networking, master-replica replication and optimistic-locking transactions (WATCH/UNWATCH).
Anchor-Med
An offline-first, P2P medical inventory system built on a custom Python B-Tree engine with WAL-based crash recovery and an Electron/React frontend.
ML & Interpretability
GlassBox-ML
An educational, NumPy-only ML library that shows why a model behaves the way it does: loss history, gradients, assumption checks, failure modes and generalization estimates. Basis of a paper in progress.
GlassBox-DL
A model-agnostic deep learning interpretability library that extends the GlassBox philosophy to neural nets: saliency maps, Grad-CAM and activation probing.
03 Blog
Notes on building things from scratch and what I learn along the way.
Building Kedis: A Redis-Inspired In-Memory Data Store From Scratch
Skip lists with rank-aware spans, a custom wire protocol (KESP), a move from threads to asyncio backed by benchmarks, AOF persistence and its durability-throughput trade-off, transactions with optimistic locking, streams, and replication. What building a database taught me about how systems fit together.
Read post →Why I Built GlassBoxML?
I wanted to understand what really happens inside model.fit(): how gradient descent updates weights, how a decision tree picks its splits. So I rebuilt the algorithms on plain NumPy. The math turned out to be the easy part; API design, validation, numerical stability, testing and packaging were the real work. “The best way to understand an abstraction is to remove it.”
04 Interests
Compilers
Lexers, parsers, IRs and code generation. The same “build it to understand it” itch, applied to languages.
Astronomy
Looking up, and wondering how we know what we know about things that far away.
ML / AI & DL
Not just training models, but opening them up: interpretability, failure modes, learning theory.
Python
My daily driver, from asyncio servers to NumPy-from-scratch learning libraries.
05 Stack
06 Contact
Want to talk systems, ML interpretability, compilers or the night sky? Get in touch.