New Delhi, India IST

Open to Summer 2027 internships

A ridge field, drawn live. Hover or tap to read its direction.

I’m a third-year engineering student at IIT Delhi. I build ML systems and signal processing tools from scratch, then test them until the numbers hold up.

See my work Email me

aaditya kumawat

Selected work5

Things I built on my own in 2026, and the numbers that came out of them.

When
Jun–Aug 2026
Kind
Self-led project
Stack
C++20, React, TypeScript, CMake

SextantOntology and lineage platform

A mini-Foundry for maritime data. It pulls port and vessel records from three public sources, merges duplicates into real entities, and can trace every stored value back to the source row and transforms that produced it. Underneath is an LSM storage engine I wrote from scratch.

of 4,201 values replay from source
100%
batched writes per second
1.53M
F1 on entity resolution
0.991
Read the case studyCode on GitHub
# cluster, fuse, write entities with provenance
$ sextant resolve
  1451 source records, 470 candidate pairs
  1187 entities written, 4201 properties,
  4201 provenance records
  dedup ratio 0.1819  (1451 records -> 1187 entities)

# the lineage round trip
$ sextant explain
lineage round trip
  1187 entities, 4201 properties
  4201 verified, 0 failed  ->  100.00%
When
Jun–Jul 2026
Kind
Self-led project
Stack
PyTorch, Triton, CUDA

nanoserveLLM inference server

An LLM inference server built without vLLM or TGI: a paged KV cache, continuous batching with chunked prefill, a fused Triton attention kernel and INT4/INT8 weight quantization. vLLM only appears as an outside yardstick.

throughput over static batching
1.65×
lower p99 time to first token
12×
tests
92
Read the case studyCode on GitHub
When
May–Jul 2026
Kind
Self-led project
Stack
Python, NumPy, SciPy

ST-BEMDDirection-adaptive 2-D EMD

A method for splitting 2-D signals like fingerprints into their component layers. It bends along the ridges instead of treating every direction alike. I tested it against four baselines that I also wrote from scratch.

lower orientation error on 100 real prints
35.7%
of 30 paired prints improved
100%
of the gain traced by ablation
78%
Read the case studyCode on GitHub
When
Jun 2026 to now
Kind
Live product
Stack
Next.js 15, TypeScript, PostgreSQL

Company HubInterview prep platform

A live site that ranks LeetCode questions by how often each company asks them, with progress tracking, spaced revision and an in-browser code editor. I run it in production and fixed the problems that only show up under real traffic.

questions indexed
3,400
companies
656
languages in the editor
9
Read the case studyVisit companyhub.funCode

Also

QRT return prediction

Predicting whether a stock rises tomorrow from anonymised market data, built to resist leakage: purged cross-validation with a 20-day embargo and a 72-day holdout that tuning never sees. A simulator that knows the true signal shows the model captures 73.7% of the edge that is actually there.LightGBM, XGBoost, Optuna

Mar 2026Code

Research2

Work done in a lab and with a professor.

When
May–Jul 2026
Kind
Research intern (remote)
Stack
Applied AI Laboratory, HEC Lausanne

A fully local pipeline that turns court audio into speaker-attributed notes. Every note links back to the moment in the recording it came from, and nothing leaves the machine. In one mode the language model can only pick sentence ids, so it has no way to invent text.

word error rate
6.9%
diarization error rate
7.9%
verbatim in extractive mode
100%
Read the case study

Towards Reliable LLM-as-a-Judge Systems

A survey of five directions in using LLMs to grade other LLMs: pairwise versus pointwise protocols, checklist assessment, MCTS reasoning judges, multilingual judges and prompt-injection security, brought together into one framework.Co-authored with Suvit Vishwakarma and Avaneesh R, advised by Prof. Amartansh Dubey.

2026Paper (PDF)

Recognition

Candidate Master on Codeforces, peak rating 1938 as codeleon, reached in 9 rated contests over six weeks

2026

Rank 170 in Codeforces Educational Round 193

2026

Letter of recommendation from the Applied AI Laboratory, HEC Lausanne

2026

All India Rank in the top 3% of about 200,000 candidates who qualified JEE Advanced

2024

99.33 percentile in JEE Main, top 0.7% of about 1.5 million candidates

2024

About

I’m in the third year of a B.Tech in Production and Industrial Engineering at IIT Delhi, graduating in 2028.

Most of what I build sits where machine learning meets systems: inference servers, storage engines and signal processing methods. I like rebuilding things from first principles to understand them, and I try to break my own numbers before anyone else does.

Outside of projects, I do competitive programming, mostly on Codeforces.

Education
B.Tech, Production and Industrial Engineering, IIT Delhi, 2024–2028
Languages
C++20, Python, TypeScript
Machine learning
PyTorch, Hugging Face Transformers, scikit-learn, LightGBM, XGBoost, Optuna, NumPy, pandas
GPU and inference
CUDA, Triton, INT4/INT8 quantization
Web and backend
Next.js, React, PostgreSQL
Tools
Git, Docker, CMake, ASan, UBSan, TSan

Contact

Have an internship, a research problem or a question about my work? Write to me.

GitHub LinkedIn Codeforces CV (PDF)