adixb

Aditya Bavadekar

CS undergrad at VIT Pune. I work on backend systems, infrastructure, and Linux tooling.

Currently building Sigmoidh, a coding-agent harness in Go. Looking for internships.

Hi, I'm Aditya. I study computer science at VIT Pune, and this is my personal site.

I mostly build backend systems and AI agents, and sometimes Linux tools.

I take part in hackathons with a team called [object Object], and play CTFs on the side. I also contribute to open-source projects, with merged pull requests in Kubeflow Trainer and the Kaggle CLI.

On my blog I write about how things work under the hood. To get in touch, use the message box below.

Projects

All projects →

Python / PyTorch / FastAPI

Dhara

Flood and debris-flow damage mapping from open Sentinel-1, Sentinel-2, DEM and OpenStreetMap data.

Built for the Multimodal AI Hackathon 2026 (Track B: Mapping Flood Damage from Space).

Demo Video GitHub

Python / FastAPI / PostgreSQL

FORTX

Agent pipeline that diagnoses failed payments and runs recovery over retries, WhatsApp, SMS and voice calls.

Built for the Razorpay AI Buildathon.

Demo Video GitHub

Python / OpenEnv / FastAPI

Agentrology

Reinforcement-learning environment that evaluates AI security agents against malware and rootkits in a Linux sandbox.

Built for the Meta PyTorch OpenEnv Hackathon.

Demo GitHub

Java / Apache Flink / Kafka

Reconcyl

Real-time transaction reconciliation across systems, built on a Kafka and Flink stream pipeline.

GitHub

Dhara

Flood and debris-flow damage mapping from open Sentinel-1, Sentinel-2, DEM and OpenStreetMap data.

Built for the Multimodal AI Hackathon 2026 (Track B: Mapping Flood Damage from Space).

Live demo Video Code

Stack

  • Python
  • PyTorch
  • FastAPI
  • React
  • ONNX Runtime
  • Rasterio
  • GeoPandas
  • DuckDB
  • AWS

Remote Sensing · SAR Processing · Flood Mapping

How it works

  1. Shows where it flooded, what was damaged and which settlements are cut off.
  2. Its own Sentinel-1 processing chain, from raw GRD files to terrain-flattened backscatter.
  3. A U-Net trained on Kuro Siwo, fused with other detectors and a physics-based filter.
  4. F1 0.84 on a held-out area, scored against Copernicus EMS mapping of the August 2026 Trishuli flood.
  5. One-page situation report in English and Nepali. Educational prototype.

Screenshots

FORTX

Agent pipeline that diagnoses failed payments and runs recovery over retries, WhatsApp, SMS and voice calls.

Built for the Razorpay AI Buildathon.

Live demo Video Code

Results

  • 24×fewer LLM diagnosis calls (95.9% cache hit rate)
  • 11-pointrecovery lift vs. a 10% holdout

Stack

  • Python
  • FastAPI
  • PostgreSQL
  • Redis
  • Next.js
  • TypeScript
  • SQLAlchemy
  • LiteLLM
  • Docker
  • AWS
  • Razorpay API
  • Sarvam AI
  • Twilio

Agentic Payment Recovery · Multi-Channel Orchestration · LLM Diagnosis

How it works

  1. Catches failed payments, subscriptions and invoices.
  2. Diagnoses each failure with an LLM, with a rule-based fallback.
  3. Recovers in bounded steps: retries, WhatsApp, SMS and Hinglish voice calls through Sarvam AI.
  4. Strict per-merchant safety limits, and a holdback group to measure the effect.

Screenshots

Agentrology

Reinforcement-learning environment that evaluates AI security agents against malware and rootkits in a Linux sandbox.

Built for the Meta PyTorch OpenEnv Hackathon.

Live demo Code

Stack

  • Python
  • OpenEnv
  • FastAPI
  • Docker
  • Linux

Sandboxing · Linux Security · Reinforcement Learning

How it works

  1. Tests security agents against self-healing rootkits and APT-style infections.
  2. The infections run in a live Linux kernel inside a sandbox.

Screenshots

Reconcyl

Real-time transaction reconciliation across systems, built on a Kafka and Flink stream pipeline.

Code

Stack

  • Java
  • Apache Flink
  • Kafka
  • Spring Boot
  • PostgreSQL
  • MinIO
  • Prometheus
  • Grafana
  • Next.js
  • Docker

Stream Processing · Fault Tolerance · State Synchronization

How it works

  1. Matches transactions across systems.
  2. Records go through a multi-stage normalisation pipeline before matching.

Achievements

Open source

All contributions →

Writing

All posts →

Education

Terminal

Type help, or boot for a real Arch Linux

Send a message

It goes straight to me. No account or email needed.

Include your email or GitHub if you’d like a reply.
Press Ctrl + Enter to send