Subhajit Pradhan
ResumeGitHub
Projects

Featured case study / Full-stack engineer

Tarka Sabha

Tarka Sabha is a multi-agent AI debate platform that coordinates configurable personas, multiple model providers, and inspectable debate workflows.

Metrics

Proof in numbers

Model boundary
Provider calls stay server-side
Core workflow
Topic, roles, turns, output
Primary design goal
Inspectable multi-agent state
Data layer
Prisma-backed workflow records

Workflow proof

Debate workflow state

The product separates debate setup, agent turns, provider calls, and generated output.

  1. Topic setup
  2. Agent roles
  3. Provider boundary
  4. Debate output

Architecture

System shape

A Next.js App Router application uses PostgreSQL and Prisma for auth-backed workflow records while a provider adapter coordinates model calls and encrypted credentials.

  1. Debate setup

    User defines the topic, debate context, and agent roles.

  2. Workflow state

    Application state tracks agent turns and generated arguments.

  3. Backend boundary

    Provider keys and model calls stay outside the browser.

  4. AI providers

    Model-specific responses are coordinated behind the same product flow.

  5. Review UI

    The interface presents generated arguments for comparison and inspection.

Proof

  • Supports configurable personas and multiple model providers behind one debate flow.
  • Uses sequential and mention-aware orchestration to control speaker turns.
  • Encrypts provider API keys at rest with AES-256-GCM.

Problem

AI debate workflows need a structured way to coordinate multiple model providers, user prompts, and credentials without mixing product logic with provider-specific details.

Users and context

Built for users who want to run and compare AI-assisted debate flows through a web interface.

Product workflow

A user defines the debate context, configures model access, starts the debate flow, and reviews generated arguments from participating AI agents.

Key engineering decisions

  • Separated provider orchestration from the interface so model integrations can change without reshaping the user flow.
  • Kept credential handling behind backend boundaries instead of exposing provider keys to the browser.
  • Structured the project around debate workflow state rather than scattered prompt experiments.

Tradeoffs and limitations

  • Supporting multiple providers increases orchestration complexity compared with a single-model implementation.

Next improvements

  • Add richer debate history views.
  • Improve response comparison and citation displays.