KabyaMithunSaha

Backend / Full-Stack Engineer · Competitive Programmer

I build the backends that carry real work — legal-document pipelines, a VM cloud for universities, an online judge — and I train on the algorithms underneath: 1,500+ problems and counting.

An interactive map of projects, records and the disciplines that connect them. Activating a marker selects it and lists its connections below the map; activating it again opens its page. Press Escape to clear the selection. A plain-text version of the map follows the figure.

Plate I. The Systems Atlas — 7 projects, 8 disciplines, 2 records.Connections are drawn only where my CV supports them.
Read the map as text

Projects

Hopper · G03
Infrastructure: Go/gRPC orchestrator managing Kubernetes pod lifecycles. · Security: Keycloak OIDC SSO with roles; Cilium zero-trust policies isolate tenants. · Data: Credit billing on a double-entry PostgreSQL ledger with advisory locks. · Async & events: NATS JetStream events; auto-termination when credits run out. · Real-time: Real-time metrics streamed over server-sent events.
CodeJudge · H06
Algorithms: An online judge: the home ground of competitive programming. · Security: Sandboxed multi-language execution with resource limits. · Infrastructure: Dockerized for scalable deployment. · Evaluation: Unit, integration and load testing. · Real-time: Real-time leaderboards. · Applied AI: AI-powered problem generation.
Photocard Forgery · C07
Applied AI: CNN and YOLOv8 classifiers reading photocards directly from pixels. · Evaluation: Leakage-free stratified group 5-fold cross-validation.
Text2SQL · E06
Applied AI: Google Gemini turns natural-language questions into SQL. · Data: PostgreSQL, 1NF–3NF normalization, pooling and query caching. · Security: SQL sanitization before anything executes. · Evaluation: 54/54 core system tests passing.
SecureVault · I05
Security: AES-256 + RSA hybrid encryption, secure sharing and access control. · Infrastructure: Containerized with Docker.
Study Helper Pro · D05
Applied AI: Retrieval-augmented answers from Google Gemini. · Async & events: PDF processing on Celery workers backed by Redis. · Data: ChromaDB embeddings with per-user isolated collections. · Security: Supabase authentication with JWT.
Hill Climb Racing · J08
Algorithms: Custom collision detection, written from scratch. · Real-time: Game-loop optimization.

Records

Makebell · document pipeline
Applied AI: Clause extraction, template generation, draft review; OCR with Qwen Vision. · Async & events: Services integrated into an asynchronous production pipeline. · Infrastructure: 3+ services on Docker Compose with health checks and resource limits.
Contest archive
Algorithms: Codeforces Expert (1672), 1,500+ problems, ICPC Asia Regional twice. · Security: CTF results: DUCTF 2024 champion, NSU Cybernauts 2026, Al Khwarizmi 2025. · Applied AI: Bengali LLM hallucination detection datathon, IUT 12th ICT Fest 2026.

Disciplines

Infrastructure
Hopper: Go/gRPC orchestrator managing Kubernetes pod lifecycles. · CodeJudge: Dockerized for scalable deployment. · SecureVault: Containerized with Docker. · Makebell · document pipeline: 3+ services on Docker Compose with health checks and resource limits.
Security
Hopper: Keycloak OIDC SSO with roles; Cilium zero-trust policies isolate tenants. · CodeJudge: Sandboxed multi-language execution with resource limits. · Text2SQL: SQL sanitization before anything executes. · SecureVault: AES-256 + RSA hybrid encryption, secure sharing and access control. · Study Helper Pro: Supabase authentication with JWT. · Contest archive: CTF results: DUCTF 2024 champion, NSU Cybernauts 2026, Al Khwarizmi 2025.
Algorithms
CodeJudge: An online judge: the home ground of competitive programming. · Hill Climb Racing: Custom collision detection, written from scratch. · Contest archive: Codeforces Expert (1672), 1,500+ problems, ICPC Asia Regional twice.
Applied AI
CodeJudge: AI-powered problem generation. · Photocard Forgery: CNN and YOLOv8 classifiers reading photocards directly from pixels. · Text2SQL: Google Gemini turns natural-language questions into SQL. · Study Helper Pro: Retrieval-augmented answers from Google Gemini. · Makebell · document pipeline: Clause extraction, template generation, draft review; OCR with Qwen Vision. · Contest archive: Bengali LLM hallucination detection datathon, IUT 12th ICT Fest 2026.
Data
Hopper: Credit billing on a double-entry PostgreSQL ledger with advisory locks. · Text2SQL: PostgreSQL, 1NF–3NF normalization, pooling and query caching. · Study Helper Pro: ChromaDB embeddings with per-user isolated collections.
Async & events
Hopper: NATS JetStream events; auto-termination when credits run out. · Study Helper Pro: PDF processing on Celery workers backed by Redis. · Makebell · document pipeline: Services integrated into an asynchronous production pipeline.
Evaluation
CodeJudge: Unit, integration and load testing. · Photocard Forgery: Leakage-free stratified group 5-fold cross-validation. · Text2SQL: 54/54 core system tests passing.
Real-time
Hopper: Real-time metrics streamed over server-sent events. · CodeJudge: Real-time leaderboards. · Hill Climb Racing: Game-loop optimization.

01 Field notes

I’m a Computer Science and Engineering undergraduate at the University of Dhaka and a backend developer at Makebell. I work where systems, algorithms, infrastructure, security and applied AI meet — and each of them makes me better at the others.

Competitive programming taught me to look for the edge case first. Capture-the-flag contests taught me to assume someone else is already looking for it. Production backends are where both lessons get tested: the document that is 100 MB, the two charges that land at the same moment, the query a model wrote that nobody should run.

Kabya Mithun Saha, smiling, in a grey suit with a blue tie.
Kabya Mithun Saha
Currently
Backend Developer, Makebell LTD — since October 2025
Studying
B.Sc. in Computer Science and Engineering, University of Dhaka — expected January 2027
Competing
Codeforces Expert (1672) · CodeChef 4★ · CTF
Organizing
General Secretary, CSEDU Informatics Club

Work, drawn to scale

Seven projects across infrastructure, applied AI, security and algorithms. Each has its own case study; three of them are where most of my thinking went.

Plate G03

Flagship · Self-hosted VM cloud platform

Hopper

A VM cloud platform for universities that slices bare-metal compute into isolated, SSH-accessible containers — with single sign-on, credit-based billing on a double-entry ledger, an event pipeline and zero-trust tenant isolation.

  • Go
  • gRPC
  • FastAPI
  • Kubernetes
  • NATS JetStream
  • PostgreSQL
  • Keycloak
Read the case study

Where the work happens

Production backends, a shipped mobile app, and three years of annotating the kind of problems I like to solve.
  1. Oct 2025Present

    Backend DeveloperMakebell LTD

    Backend services for AI-powered legal document processing: extraction, generation, review and comparison.

    • Architected FastAPI microservices for AI-powered legal document processing: clause extraction, template generation and draft review.
    • Built a DOCX comparison engine that produces redline reports for documents up to 100 MB.
    • Built an OCR microservice with Qwen Vision and PyMuPDF.
    • Integrated these services into an asynchronous pipeline for production workloads.
    • Containerized 3+ services with Docker Compose, with health checks, resource limits and separate development and production configurations.
    • FastAPI
    • Python
    • Qwen Vision
    • PyMuPDF
    • Docker Compose
  2. May 2025Aug 2025

    Android Developer (Flutter)Sociofi Technology

    A Flutter app for an internal employee management system, used by production teams day to day.

    • Built and shipped an internal employee management app in Flutter, used by production teams for daily operations.
    • Implemented role-based access control.
    • Worked on the APIs behind it, with request validation and authentication.
    • Delivered production-ready features under tight deadlines.
    • Flutter
    • Dart
    • Android
    • REST APIs
  3. Aug 2023Present

    Programming AnnotatorGenMorphics AI Solutions

    Python and mathematical problems, annotated for AI training datasets.

    • Annotated Python and mathematical problems for AI training datasets, with attention to correctness, edge-case coverage and validation consistency.
    • Python
    • Mathematics

Fig. 3.1 — Document pipeline

Conceptual

  1. 01Document inDOCX · PDF · scans
  2. 02OCR serviceQwen Vision · PyMuPDF
  3. 03Clause extractionAI-powered
  4. 04Template generationAI-powered
  5. 05Draft reviewAI-powered
  6. 06DOCX comparisonup to 100 MB
  7. 07Redline reportoutput
Illustration. The stages named on my CV for my work at Makebell — not the company’s actual architecture, code or data.

The record, entered in ink

Algorithms, capture-the-flag and datathons. Ratings are the values on my CV, not live data; years appear only where the CV gives them.
Codeforces Expert (opens in a new tab)
1672
Rating as listed on my CV
CodeChef (opens in a new tab)
4★
Rating band as listed on my CV
Problems solved
1,500+
Across competitive programming judges
ICPC Asia Regional
2×
Participated twice
National contests
15+
Participated
Competition results by year
YearEventDisciplineResult
2026NSU CybernautsCTF1st Runner-up
IUT 12th ICT FestDatathon: Bengali LLM hallucination detectionDatathon2nd Runner-up
2025Al Khwarizmi CTFCTF2nd Place
2024DUCTFCTFChampion
Battle of BrainsProgrammingRunner-up
2023Battle of BrainsProgrammingRunner-up
UndatedCUET DevOps HackathonHackathon4th Place

05 Correspondence

Write to me

Backend and full-stack roles, research collaborations, contest organizing — or a system you think I’d enjoy taking apart. Email is the best way to reach me.

Site index