Featured
14 August 2026 · 3 min
Why I keep ending up between solvers and models
Scientific machine learning, LLM safety, and low-latency databases look like three unrelated fields. Working across them, they rhyme more than they differ.
Data Science & AI @ NTU
Aryan Dutt
Data Science & AI undergraduate at NTU Singapore, working across low-latency infrastructure, scientific machine learning, LLM safety, and quantitative systems.
Work locations
10 roles across three cities.
- Singapore: NTU (On-site), InterSystems (On-site), NTU CCDS (On-site), NTU SC3DP (Hybrid), SEAQR / SeaSwarm (On-site), Tagit (On-site), BioMetallica (On-site), Ministry of Home Affairs (On-site)
- Cambridge, MA: MIT Julia Lab (Remote and on-site)
- London: General Learning (YC F24) (Remote)
On-site and hybridRemote
Affiliations
- NTU
- MIT Julia Lab
- InterSystems
- YC F24
- Citadel Securities Discover
Profile at a glance
- Focus
- Systems / ML / Quant
- Base
- Singapore
- Languages
- Python, C++, Julia, TypeScript
- Current interests
- Market microstructure, evaluation harness design, and systems where tail latency is the metric that matters
- Status
- Building and researching, open to 2027 roles and collaborations
01/Selected work
All projectsQuant
Sep 2026 – Present
C++ Limit Order Book with Latency Instrumentation
A matching engine that treats latency as a measured quantity rather than an assumption.
- C++17
- CMake
- Market Microstructure
- Latency Analysis
Research
Jan 2026 – Mar 2026
Empirical Study of RoA Penalty Functions in NeuralLyapunov.jl
A controlled study of how region-of-attraction penalties shape stability certificates learned by physics-informed networks.
- Julia
- NeuralLyapunov.jl
- NeuralPDE.jl
- Lux.jl
- ModelingToolkit.jl
ML
Aug 2025 – Dec 2025
Federated Whisper Aggregation Pipeline
Fine-tuning a production speech model across decentralised datasets without moving any data.
- Python
- Flower
- PyTorch
- HuggingFace
- Whisper
ML
Apr 2025 – Jun 2025
MoE Compression of LLM Weights
Post-training compression of Mixture-of-Experts checkpoints, benchmarked for semantic drift rather than size alone.
- PyTorch
- Quantization
- SVD
- Model Compression
Quant
Nov 2024 – Jan 2025
Portfolio Optimization Engine
A multi-asset risk and simulation engine built to production shape, not notebook shape.
- Python
- NumPy
- Monte Carlo
- pytest
- Tkinter
02/Experience
Full timelineResearch
MIT Julia Lab
Research Intern
Researched scientific machine learning in the Julia Lab under Dr. Chris Rackauckas, working where numerical solvers and learned models have to agree with each other.
Jul 2025 – Mar 2026
Engineering
InterSystems
Software Engineer Intern
Engineered low-latency platform work behind real-time database workloads, building across the ObjectScript stack with CMake and Ansible so environments stayed reproducible.
May – Jul 2026
Research
NTU CCDS
Research Intern
Co-authored a paper on LLM safety alignment, building the evaluation harness that grouped failures into prompt families so the results said something the headline refusal rate could not.
Aug 2025 – May 2026
Engineering
Tagit
Software Engineer
Architected a modular transaction anomaly detection system for digital banking, modelling outliers with per-user Isolation Forests and shipping the dashboard that let analysts act on what it flagged.
Jan – Mar 2025
Engineering
General Learning (YC F24)
Intern
Built the Node.js scraping and taxonomy infrastructure behind A Level and GMAT curriculum data, normalising scraped syllabi into a hierarchy the backend could actually query.
Aug – Oct 2024
Service
Ministry of Home Affairs, Singapore
National Service, ProCom
Served two years under the Protective Security Command on critical infrastructure protection, in an environment where operational discipline was the whole job.
Nov 2022 – Oct 2024
03/Research notes
All writing04/Contact
Building something difficult?
I'm interested in research, systems, and quantitative engineering problems where correctness and performance both matter, and where the answer has to hold up under load rather than only in the notebook.