Kusha Sahu
Senior Software Engineer, Raft AI
Founder, Starithm · she/her

Distributed systems, data platforms, and the occasional supernova.

Nine years building backends, ML pipelines and geospatial products that serve more than 10 million users. I write about the papers behind them, and I'm building Starithm to turn real-time astronomy alerts into structured intelligence.

Now / Day job

Raft AI

Scalable AI infrastructure and data-intensive systems as an IC-L4, owning architecture for core engineering initiatives.

Now / Building

NovaTrace ingests multi-messenger alerts in real time. React microfrontends, TypeScript APIs, Python ingestion, Postgres, Kafka.

Now / Writing

A thought journal on Substack. Everyday thoughts from everyday monotony, written down as they come.

What I work with

Grouped by the layer I actually spend time in, rather than every tool I have ever opened once.

Distributed systems & backend
KafkaPostgreSQLTypeScript APIsPythonEvent-driven ingestionReal-time streamingETL at scale
ML & data
Transformer ensemblesCNN-LSTMXGBoostTime-series modellingImbalanced data (SMOTE)LangGraphRetrieval pipelines
Product & frontend infra
React microfrontendsInternal TS framework (Orion)Micro-frontend monitoringGeospatial maps productsSubscription systems
Cloud & operations
AWS LambdaS3RekognitionCDN cost optimisationAutomated monitoringTeam leadership

Work

Nine years, five teams, one throughline: systems that stay correct at scale.

Raft AI

Senior Software Engineer (IC-L4)
Feb 2026 - Present

Building scalable AI infrastructure and data-intensive systems. Driving architectural decisions and execution for core engineering initiatives.

Starithm

Founder & Developer
Aug 2025 - Present

AI-powered astronomy event intelligence platform bridging the gap between real-time multi-messenger astronomy alerts and actionable structured intelligence.

  • Built NovaTrace (MVP) to aggregate and process astronomical alerts.
  • Architected React microfrontends, TypeScript APIs, Python ingestion, PostgreSQL, Kafka.
  • Engineered real-time streaming ingestion and event-based aggregation logic.

Aurora Energy Research

Senior Software Engineer / Head of Indian Software Team
Sept 2024 - Jan 2026
  • Owned and scaled a complex geospatial maps product used for energy infrastructure analysis.
  • Architected internal Gen-AI Knowledge Chatbot from scratch using LangGraph and AWS Lambda.
  • Optimized historical and real-time energy ETL workflows.

Urban Company

SDE 1 → SDE 2 → SDE 3
2018 - Aug 2024
  • Designed automated monitoring for 10+ micro-frontends.
  • Built image monitoring pipeline saving 2TB/day in CDN transfers.
  • Designed subscription product contributing 23% of revenue.
  • Architected internal TypeScript-based frontend framework (Orion).

Tata Digital Health (Tata 1MG)

Associate Programmer
2017 - 2018
  • Built pharmacy support dashboards and implemented secure S3 file uploads.
  • Built face-recognition API using AWS Rekognition.
  • Optimized SEO search algorithms.

Projects

Astronomy, ML and large-scale data. Mostly things I started because a paper made me curious at 1am.

01

Starithm — NovaTrace, real-time multi-messenger alert intelligence

Ingests the NASA GCN alert stream in real time and maintains one evolving record per astrophysical event, stitched from machine notices and human-written circulars.

  • Kafka consumer plus a local LLM pipeline that extracts measurements out of circular prose.
  • Event identity resolved through an explicit per-source registry, replacing a heuristic that left ~11% duplicate records.
  • 5,050 notices across 2,057 events; p50 ingestion latency 1.0s from GCN publication to stored record.
02

Retro Agent: analytics for coding agent transcripts

Reads what Claude Code, Codex and Cursor already write to disk and answers two questions their own dashboards do not: where you had to repeat yourself, and whether a session is safe to show anyone else. Local only, no network calls.

  • One representation across three agents, keyed on what a tool call did rather than what it was named, so the same metrics hold whether an agent names its tools or pushes everything through a shell.
  • Rework measured from transcript structure instead of token volume: verification runs per edit, files edited repeatedly with no test in between, byte-identical calls re-issued. Across 163 sessions, cached context replay outnumbered generated tokens 296 to 1.
  • Credentials stripped before anything reaches the archive. 366 connection strings and every JWT down to zero, checked by re-running the scan that found them.
  • A parity command that refuses to compare two agents on a metric neither adapter can see properly, because a blind spot otherwise looks exactly like a behavioural difference.
03

Space Debris Detection and Trajectory Prediction

Built a two-part system for orbital debris management using ML classification and trajectory modeling.

  • Applied SMOTE balancing for rare debris classes in detection models.
  • Modeled 5.8M+ orbital time-series entries to predict position and velocity vectors.
04

Measuring Hate Speech Exposure Across Algorithmic Feeds on Bluesky

Developed an ensemble moderation auditing pipeline to evaluate toxicity exposure across algorithmic feeds on Bluesky.

  • Combined Detoxify, CardiffNLP, Dynabench transformer classifiers with Mistral LLM.
  • Introduced new metrics: Toxicity Leak Score and Feed Toxicity Divergence.
  • Analyzed 150k+ posts, identifying toxicity leakage disparities up to 10%.
05

Exoplanet Detection using Machine Learning

Designed ML pipeline for transit detection from light curves.

  • Implemented CNN-LSTM architecture and custom transit dip feature engineering.
  • Conducted temporal-spectral modeling and XGBoost baseline comparisons.
06

Kaggle Starters

Starter notebooks and experiments on Kaggle.

07

Activity Recognition using MHI and Hu Moments

Video classification using Motion History Images and Hu Moments with ML classifiers.

08

Solving Raven's Progressive Matrices using KBAI

Knowledge-based AI approach to solving Raven's Progressive Matrices.

09

Crossmatching Catalogs using kd Trees (Nearest Neighbor)

Crossmatching astronomical catalogs using kd-tree nearest neighbor search.

10

Median Image Stacking – Binapprox Algorithm – Detection of Pulsars

Increasing S/N ratio via median image stacking and Binapprox algorithm for pulsar detection.

Deep dives on scalable backends, ML architecture and turning unstructured data into structured intelligence. Plus 2 Cents at Coreforge, where I think out loud about ordinary days.

Latest

Conversations with Codents: Episode 2 — Where Agent Disagreed With Itself

Read on Medium ↗

Reading list

Yes, I have read these. Most of them. The ones more than once, the * ones only where a library let me in.

Books [Only My Favs]17

  • Weapons of Math Destruction – Cathy O'Neil
  • The Brain – David Eagleman
  • QED – Richard Feynman
  • The Ascent of Gravity – Marcus Chown
  • Genius Makers
  • MIT Press: AI Ethics
  • Deep Work – Cal Newport
  • Emotional Intelligence
  • A Brief History of Time – Stephen Hawking
  • Work Rules
  • The Psychology of Money
  • Soonish
  • The Brain
  • Notes on a Nervous Planet
  • QED – Richard Feynman
  • Our Mathematical Universe
  • Sapiens: A Brief History of Humankind

Tech Books8

  • Designing Data Intensive Applications
  • Designing Machine Learning Systems
  • Machine Learning Design Patterns
  • Fundamentals of Data Engineering
  • Clean Code
  • Structure and Interpretation of Computer Programs
  • You Don't Know JS
  • Hundred Page Machine Learning Book

Want to Read1

  • Crucial Conversations – Patterson, Grenny, McMillan, Switzler