Skip to content

My work

Things I have built

AI agents, language models, computer vision and the products around them. Filter by area, or open any project on GitHub.

7 projects shown

  • AI systems · Node.js

    SuprBuild Mesh

    An agent task mesh for publishing, claiming, submitting, verifying and settling work. Built with Node.js, Express and SQLite, with reputation scoring, wallet flows and a live dashboard.

    Read case study
    • 20+ REST endpoints
    • 10+ concurrent agents
    • ~120 ms median response (local load tests)
    • Node.js
    • Express
    • SQLite
    • REST
    View on GitHub ↗
  • Bangla AI · TypeScript

    Route52

    A Bangla-native TypeScript and React assistant with authentication, persistent conversations, agent-oriented workflows, rich Markdown, and readable code blocks.

    Read case study
    • Beta-tested with 30+ users
    • 200+ conversations used to refine prompts
    • Persistent chat history
    • TypeScript
    • React
    • LLMs
    • Authentication
    View on GitHub ↗
  • Fintech · TypeScript

    In progress

    Krostio

    A full-stack income-verification product concept for gig workers and freelancers. It turns connected earnings into income histories, consistency insights, projections and shareable lender-ready reports.

    Read case study
    • Income verification reports and consistency scoring
    • Plaid integration is configured for sandbox data
    • PDF reports and optional on-chain attestations
    • Next.js
    • TypeScript
    • Supabase
    • Plaid Sandbox
    • Paddle
    • Base
    View on GitHub ↗
  • Research reproduction · PyTorch

    PhyloVAE

    A PyTorch implementation and experiment based on published PhyloVAE research for learning representations of phylogenetic tree topologies.

    Read case study
    • CV-reported reconstruction loss reduction of ~18% vs a baseline autoencoder
    • Based on an ICLR 2025 paper by Xie et al.
    • Python
    • PyTorch
    • Variational autoencoders
    • Phylogenetics
    View on GitHub ↗
  • Computer vision

    Quick-Draw Detection

    A QuickDraw sketch-classification experiment comparing CNN and Bi-LSTM models. The CV reports about 92% test accuracy, six percentage points above the CNN-only baseline.

    Read case study
    • ~92% test accuracy (CV-reported)
    • 6 percentage points above the CNN-only baseline
    • Python
    • Computer vision
    • CNN
    • Bi-LSTM
    View on GitHub ↗
  • AI training · Next.js

    In progress

    Interviewer Practice Lab

    AI-simulated candidates that let hiring managers practise interviewing before they sit across from a real person.

    • Synthetic candidates for interviewer training
    • Next.js
    • LLMs
    • SaaS
  • EdTech · Next.js

    In progress

    Admission Mock Test Builder

    An AI mock-test builder and weakness finder for BUET-style and DU admission exams.

    • Generates exam-style mock tests
    • Finds each student's weak topics
    • Next.js
    • Tailwind
    • Supabase
    • OpenAI-compatible API