I build AI agents that run real lending workflows.

Featured · Desktop app · v1.1.0 for Windows & macOS

Ascend: an AI interview coach built on the Claude Agent SDK

A Duolingo-style coach with no pre-written course. Claude designs a 10-level roadmap for you, writes each lesson, grades your answers like an interviewer, runs mock interviews and researches the company you're interviewing with. It adapts to what you already know.

1. Learner stateMCP tools
2. Claudejson_schema output
3. Validate & repairguardrail
4. Mastery engineElo + SM-2
5. Next lesson

Agent engineering

• Least-privilege sessions. Every agent runs with a strict MCP config and gets only the tools its task needs: WebSearch, WebFetch and the app's own tools. It has no shell, no file access, and no user hooks or CLAUDE.md.
• In-process MCP server with six tools: get_learner_profile, get_topic_mastery, get_recent_mistakes, record_coach_note, suggest_practice, schedule_review.
• Schema-enforced output. Roadmaps, questions, grades and interview scorecards are JSON-schema outputs. They are validated and retried, and every question is repaired if needed before it renders.
• Long-lived sessions. The mock interviewer and the coach chat keep one agent process alive across turns using streaming input, and they resume after idling or a restart.
• Latency work. The grader process starts warming up while you are still typing. Primers stream token by token and show live tool activity.
• Memory the user can see. After each lesson the agent writes a debrief and saves lasting observations. It can add a remedial topic to your path, and you can edit all of it.
• A deterministic core. Claude writes the lessons and grades the answers. An Elo-style mastery score and SM-2 spaced repetition decide what comes next.

• Claude Agent SDK
• MCP
• TypeScript
• Electron
• React 19
• Tailwind 4

Visit site Source Download

Chrome extension · Manifest V3 · MIT

TL;DR: an assistant for any web page that uses MCP tools with your approval

It summarizes any web page or PDF and lets you chat with it. If you connect remote MCP servers, the model can act on what you're reading. For example, it can log the expense from a bank alert or check the ticket a page mentions, but only after you approve the call.

1. Page text
2. ModelGemini / Groq
3. Tool call?
4. Approval cardRun · Skip
5. MCP server
6. Grounded answer

Agent engineering

• MCP client written by hand. It speaks JSON-RPC over the Streamable HTTP transport, with a handshake, tool discovery and auth headers. No SDK, no build step.
• A human approves every call. Each tool call shows the tool, the reason and the exact arguments. If you press Skip, the model is told and answers with what it already has. An unanswered prompt counts as Skip after two minutes.
• Limits on runaway loops. Each message has a cap on tool calls, and closing the popup declines any call that is still pending.
• Quick actions written by the model. After a summary, the model suggests one-tap actions for that page, such as Log ₹4,800 expense. Actions linked to tools run when tapped, and the tap counts as approval.
• Tools are used only where they're needed. Tools are offered only in Q&A. A summary is always a single fast call.
• Works with more than one provider and keeps data private. It runs on Gemini or Groq (Llama 3.3, GPT-OSS), and you choose the model. Your keys stay in the browser, PDFs are parsed locally, and there is no backend or telemetry.

• MCP client
• Tool calling
• Gemini
• Groq
• Vanilla JS
• Chrome MV3

Source & install

More on github.com/kaushikappani, including NoteIt (React, Node, MongoDB, Redis) and earlier experiments.

1. 

   Rahi Platform Technology

   Pune

   Senior Software Engineer · AI & Agentic Systems

   Jul 2026 – Present

   • Leading development of an agentic loan origination system that uses LLM-based orchestration across the full loan processing workflow.
   • Designed an MCP framework with LLM pipelines that analyzes collection call recordings, scores call quality and sentiment, and automates the workflow decisions that follow.
   • Built an AI voice tele-verification agent that asks context-aware, dynamic questions, which reduces manual verification work.
   • Run production agents on AWS Bedrock with LangFuse tracing to improve performance, observability and reliability.

   Software Engineer · Loan Management System

   Jul 2025 – Jul 2026

   • Delivered a cloud-agnostic, multi-tenant LMS as SaaS, covering collections, payments, accounting, reconciliation and customer communication.
   • Built orchestration engines for payment appropriation, collections allocation, penalties, reversals and accounting entries.
   • Introduced an event-driven architecture on Kafka that connects the accounting, collections, notification and workflow services.
   • Built Spring Batch end-of-day jobs with retry and checkpointing that process 30,000+ loan accounts in 15 minutes.
   • Integrated payment gateways, dialer systems and SMS, email and WhatsApp channels for automated collections.
2. 

   Bajaj Finserv Direct (Bajaj Markets)

   Pune

   Software Engineer II · Bajaj AMC Distributor Portal

   Jun 2024 – Jul 2025

   • Built backend services for a high-traffic platform used for mutual fund distributor onboarding, partner management and financial operations.
   • Designed role-based access control (RBAC) and authorization workflows for a user hierarchy of distributors, admins and operations teams.
   • Integrated Kotak APIs for real-time KYC, account verification and onboarding.
   • Cut response times through query, indexing and caching work on PostgreSQL and MongoDB.

   Software Engineer I · UGRO Capital credit CRM

   Jun 2023 – Jun 2024

   • Built backend modules for a credit and loan-processing CRM: onboarding, lead workflows, loan processing and collections.
   • Automated workflows and business-rule validation to cut manual work and turnaround time in credit operations.

   Byte Intern · Enterprise Applications

   Jan 2023 – Jun 2023

   • Built REST APIs and backend logic with Java and Spring Boot, and helped debug production issues.