aarin.basu
Open to software engineering roles — infrastructure, backend & applied ML

Aarin
Basu

I build backend, ML, and systems software — from edge-deployed AI to a vector database written from scratch in Rust.

View work ↓GitHub ↗LinkedIn ↗Download résumé ↓
·

built with

Python
Rust
PyTorch
FastAPI
React
TypeScript
Docker
AWS
PostgreSQL
Python
Rust
PyTorch
FastAPI
React
TypeScript
Docker
AWS
PostgreSQL
01 —

Featured Projects

tidemarkIn progress

Exactly-Once Stream Processing in Go

  • Go
  • Pebble
  • Stream Processing
  • Nexmark
  • Event-time stream processing engine in Go: watermarks propagate through the dataflow graph as the minimum across input channels, so a window fires only after every upstream partition has passed it.
  • Chandy-Lamport barrier checkpointing with Pebble-backed operator state. Sinks commit only once a checkpoint is confirmed, so a run that fails mid-job and recovers ends with the same sink contents as a clean run.
  • Faults are injected at deterministic logical positions, never on a wall clock, and a chaos suite runs seeded fault schedules across five Nexmark queries with windowed results checked against a batch oracle.
See how it works →Repository ↗Benchmarks ↗

flakescopeOpen source

Flaky Test Detector for Go

  • Go
  • CLI
  • Testing
  • CLI that reruns a Go package's tests across a matrix of configurations (test order, GOMAXPROCS, race detector) and reports which tests fail nondeterministically, with the minimal configuration that reproduces each.
  • Names the knob a failure depends on by comparing failure rates across axes with a two-proportion z test, and reports undetermined instead of guessing when a run is too small to resolve the rate.
  • Clusters failures by normalized signature: one 1,000-configuration run turned 2,174 individual failures into four findings. Zero non-stdlib dependencies, enforced by a test.
See how it works →Repository ↗

AEGISFlagship

AI-Enhanced Combat Triage Copilot

  • Python
  • PyTorch
  • YOLOv8
  • Whisper
  • Streamlit
  • Jetson
  • Edge-deployable copilot fusing real-time computer vision and audio for mass-casualty triage — running fully offline on NVIDIA Jetson at 15–30 FPS.
  • Multi-model perception pipeline: YOLOv8 detection, ByteTrack/DINOv2 re-identification, MobileSAM wound segmentation, and zero-shot CLAP audio for respiratory monitoring.
  • Streamlit tactical dashboard with Whisper voice commands, rule-based SALT/TCCC decision support with human-in-the-loop confirmation, and automated 9-line MEDEVAC generation.
See how it works →Repository ↗

vexdbLive demo

A Vector Database in Rust

  • Rust
  • HNSW
  • SIMD
  • WASM
  • axum
  • proptest
  • A vector database written from scratch in Rust with a hand-rolled HNSW index, AVX2 distance kernels, binary snapshot persistence, traversal-time payload filtering, and a Qdrant-shaped HTTP API. No faiss bindings.
  • Benchmarked head to head against faiss on identical data: recall curves within ±0.03 at every ef, with the remaining 2 to 2.5× QPS gap isolated to distance-kernel throughput rather than graph quality.
  • The whole engine compiles to a 289 KB WebAssembly module. The live demo searches 5,000 movie-plot embeddings in your browser and animates the engine's real traversal trace.
See how it works →Live demo ↗Repository ↗

limitbookLive demo

ITCH 5.0 Parser & Limit Order Book Engine

  • Rust
  • zero-copy
  • no_std
  • WASM
  • criterion
  • CI
  • A zero-copy parser for the NASDAQ TotalView-ITCH 5.0 binary market-data feed, reconstructing the full limit order book across a 268.7M-message trading day with zero invariant violations.
  • Single core, warm, from decompressed bytes: 31.9M msg/s parse-only, 1.38M msg/s with full book reconstruction. The book dominates, not the parser, and the flamegraph proves it.
  • Profiling found one crossed-book re-arming check burning 23% of the pass on work that was a provable no-op. One early return took it to 7.5%, with byte-identical replay output verified against the pre-change baseline.
See how it works →Live demo ↗Repository ↗

OutlierQInteractive demo

Options Trading Signal Engine

  • Python
  • FastAPI
  • React
  • TypeScript
  • yfinance
  • VADER
  • End-to-end signals platform processing real-time news and market data to generate options recommendations with strike-price and expiry logic.
  • Multi-stage anomaly-detection pipeline: z-score volume analysis, sentiment filtering, and keyword-based event classification.
  • Full-stack React/TypeScript + FastAPI dashboard visualizing live signals and model performance, backed by 155+ passing tests.
See how it works →Project page ↗Interactive demo ↗Repository ↗

PostPilotLive

Automated Social Posting Platform

  • React
  • Node.js
  • Google APIs
  • Converts organizational Google Calendar events into scheduled social-media posts automatically.
  • Google Calendar API integration extracting event metadata and dynamically generating formatted content.
  • Scalable backend workflows managing scheduling, formatting, and publishing across platforms.
See how it works →Launch app ↗Repository ↗
02 —

Experience

May 2026 – Present

Parabole.ai

Software Engineering Intern, Model Evaluation & Benchmarking

Built an evaluation harness benchmarking 12 models across 6 task suites on accuracy, latency, and cost, with every run reproducible. Rewrote the sweep around a concurrent worker pool that handles per-provider rate limiting, idempotent retries, and mid-run worker failure without losing completed work, taking a full 12-model sweep from roughly four hours to under 20 minutes. Added automated regression detection on per-model latency and accuracy, after a 30% latency regression surfaced during model selection and changed which model the team shipped.

Aug 2024 – Dec 2025

Sustainability Analytics · UMD

Undergraduate Researcher

Built Python pipelines (Pandas, NumPy) to clean, process, and run anomaly detection over 12K+ water-quality measurements from the Paint Branch watershed: turbidity, pH, dissolved oxygen, nutrients. Turned the seasonal and land-use trends that fell out of it into findings an interdisciplinary team could act on, with reproducible Matplotlib and Seaborn visualizations behind every chart.

May – Aug 2025

Parabole.ai

Software Engineering Intern

Owned the service layer between the model pipeline and Parabole's analytics product, shipping six FastAPI endpoints backing three downstream services in production. Profiling showed feature enrichment was 60% of request time, so I moved it to async pre-compute and batched the per-record model calls, cutting p99 inference latency from 920ms to 240ms.

Summers 2023 and 2024

City of Stamford

Software & Operations Intern

Built an internal web application consolidating 10K+ municipal records over Oracle SQL and PostgreSQL, with secure REST APIs and tuned queries behind the analytics dashboards, cutting query latency by 30%. Automated the Oracle reporting pipelines for Permitting and Licensing, taking 40% of the manual data work out of the process.

Jun – Sep 2023

University of Connecticut & State Street

Software Engineering Intern

Built a financial analytics pipeline combining REST APIs, cloud NLP, and sentiment models. Automated ingestion and preprocessing of market data in Python and SQL to 3× throughput, and built dashboards examining how market performance tracked against sentiment signals.

03 —

Live from GitHub

live from github.com/AarinB1

~/portfolio
/api/whoami · JSON
$ curl https://aarinbasu.com/api/whoami
a tiny JSON endpoint — name, focus, links, résumé URL
04 —

About

I'm a Computer Science and Mathematics student at the University of Maryland, drawn to the parts of engineering where things actually get hard — low-latency APIs, ML inference pipelines, and systems-level work like writing a database index by hand.

I like building things that have to work under real constraints: offline on a Jetson, fast enough to trade on, correct enough to ship. And I care about explaining why a system is built the way it is, not just that it runs.

05 —

Common Questions

Software engineering roles in infrastructure, backend, and applied ML — anywhere the hard parts are systems-level: low-latency APIs, inference pipelines, distributed storage, or data-intensive services. I care more about the problem than the title.

Available for summer 2027 internships full-time, and part-time during the academic year. Graduating May 2028 from the University of Maryland (CS + Math).

Based in Stamford, CT / College Park, MD during the school year. Open to relocation for internships and full-time roles in any major US tech hub, and to fully-remote teams.

Authorized to work in the US without sponsorship.

Email is the fastest path — aarinbasu1@gmail.com. LinkedIn DMs also work. Happy to share my résumé on request.
06 — LET'S TALK

Open to software engineering roles — infrastructure, backend & applied ML.

Reach me at aarinbasu1@gmail.com

GitHub ↗LinkedIn ↗Download résumé ↓