Krishnendu Biswas
Krishnendu Biswas

Krishnendu Biswas

AI agent builder 2nd-year MBA student, IIM Ranchi BS Data Science, IIT Madras

I build AI agent systems and ship the apps around them.

Agents handle my day. Every send, delete and public push waits for my approval.

A 12 minute read, or jump to any chapter.

Daily rhythm (46 active days)
Hours with agents by day
Total hours331.6 h
Active days46
Daily average7.2 h/day
Longest day13.5 h (15 Oct 2026)

A session ends after 15 minutes with no activity. Counts time on Discord, Telegram, the Hermes command line and Claude Code.

Source: timestamps of my own messages to agents, 15 minute idle gap, 5 minute tail. As of 2026-11-01.

Full operating rhythm breakdown →
Turn volume & human leverage
My messages (dictated or typed)44,023
Agent replies230,386
Tool executions232,595

About 5 agent messages for every message I sent, counting scheduled jobs.

Counted from every agent's conversation database, duplicates removed. How I count →

Source: assistant messages in every agent's conversation database, deduplicated. As of 2026-11-01.

See human approval gates →
Estate topology (8 domain agents)
  • Academics
  • Career prep
  • Finance
  • Infrastructure
  • Personal ops
  • Institute work, private
  • Socials
  • Consulting

57 recurring jobs run on the Hermes built-in scheduler.

Source: the 9 agent folders on the server. As of 2026-11-01.

Inspect platform topology →

Hover or tap any card for daily heatmaps and system breakdowns

Hours, tokens and messages are measured on the server at every build. Last build 2026-11-01. How these are counted →

Resume (PDF) See what I shipped

Email · LinkedIn · GitHub

I build with agents

I am 23 and I study management at IIM Ranchi, in Ranchi. This page shows how a management student ended up running a team of AI agents.

Chapter 2 of 10

Start here: what you can use today

I am a management student who learned to code. I built a small team of AI agents that cannot act without my approval, then used that team to ship apps for other students.

Termly Today screen on a phone: a running class with a countdown, its room and what is next (demo account)

Public beta

Termly

Term planner for IIM Ranchi MBA students. Timetable, absences, deadlines.

Open the welcome page
CogPrep Arcade Block Slide puzzle on a phone: a four by four board with a ball, three coloured blocks and the exit at the top right

Free to try

CogPrep Arcade

Six free aptitude puzzle games with a move-by-move coach.

Play now

My own tool

Agent platform

Nine agents that research and draft. I approve every outward action.

How it works

3 products, swipe

Here is how I got here

Chapter 3 of 10

Where I started

A management degree, a programming diploma and three internships.

Study, internships and positions on one time axis, from Oct 2022

solid: completed or current hatched: planned dotted: month not recorded diamond: completed on a date, month not recorded dashed line: today

  1. IPM at IIM Ranchi Oct 2022 to Mar 2027

  2. IIT Madras: BS in Data Science and Applications May 2023, in progress

  3. Senior Executive Member Executive Member from around Jul 2025, Senior Executive Member from 2026

  4. IIM Ranchi, Social Impact Intern Jan 2024 to Feb 2024

  5. Tata Steel, Research Analyst Intern Apr 2025 to Jun 2025

  6. QualityKiosk Technologies, Data and AI Intern Apr 2026 to Jun 2026

QualityKiosk internship, Apr to Jun 2026. Internal QA tooling for banking software; the code is private.

Scale

  • 18,456QA rows standardised across 61 workbooks
  • 1,027lines in 2 TypeScript estimation engines
  • 7 appsplus 3 standalone tools

Quality

  • 136unit tests, 68 of them check parity with the Excel model

Spec

  • 1,147line ML estimation PRD: 6 regression targets, 10 features, champion challenger rules

Delivery

  • 144commits, solo, in 7 weeks
Earlier internships

IIM Ranchi

  • 390+rural students assessed with Pratham and Unnat Bharat Abhiyan

Tata Steel

  • 20+manufacturing firms in a regional pay gap analysis
  • 50+ job profilescompared in the same analysis
Credentials and results, oldest first
  • 2023IIT Madras Foundational LevelProgramming and Data Science
  • 2025Topper Badges, IIT Madras7 of the first 14 courses
  • 2025Lean Six Sigma Green BeltGrant Thornton
  • 2025Digital Supply Chain TwinsanyLogistix, IIM Mumbai
  • 2025Microsoft Power BI Data Analyst (PL-300)passed Jun 2025
  • 2025Odoo x Paradox hackathontop 50 of 500 teams

All internship details

Then I learned the models

Chapter 4 of 10

Learning the models first

Modelling projects and competitions came before any agent was allowed to act.

Comment classification: F1-Macro, baseline to blend
Logistic Regression 0.6221 LightGBM 0.8202 3-model blend 0.8293 +0.198 +0.009

Zero based. 198K comments, 4 imbalanced classes, so F1-Macro rather than accuracy.

Customer exit: validation F1 of 8 classifiers, then the tuned winner
Tuned Random Forest 0.6365 XGBoost 0.6277 LightGBM 0.6277 Gradient Boosting 0.6006 Random Forest 0.5796 Extra Trees 0.5736 Logistic Regression 0.5509 Decision Tree 0.5033 AdaBoost 0.4604

Zero based. XGBoost and LightGBM tied at 0.6277. Accuracy was 0.8414 while F1 was 0.6365: few customers leave, so I report F1.

Smart MCQ solver: how the two models are blended (weights)

0.74242 MAP@3 · Kaggle public leaderboard

Both models fine-tuned with LoRA (rank 8, alpha 16).

Consumer segmentation: 62 respondents, silhouette
  • Kept: 5 clusters on 2 spend dimensionsk=5 · 0.540

Other feature sets, best k each, not kept

  • 2 spend dimensions, unscaledk=7 · 0.564
  • 2 spend dimensions, scaledk=8 · 0.529
  • 3 spend items, scaledk=2 · 0.433
  • 3 spend items, unscaledk=8 · 0.418
  • 3 spend items plus age, scaledk=2 · 0.351

I chose 5 clusters from the elbow on within-cluster spread and confirmed it with silhouette 0.540. Adding more spend or age columns scored lower, so I kept 2 spend dimensions. All bars share one 0 to 1 scale.

More builds, each with its headline fact

  • 7 DAX measures

    Venture capital deal dashboard

    7 DAX measures on a 50,000 row synthetic schema, so no client data is exposed.

  • 12 DAX measures

    HR attrition dashboard

    2-page Power BI, hires against the prior quarter.

  • FastAPI + embeddings

    Virtual teaching assistant

    Course content chunked, embedded and retrieved by cosine similarity.

  • Flask, 10+ dashboards

    IESCP platform

    Influencer and sponsor platform, deployed for 50+ influencers, sponsors.

Three IIT Madras machine learning assignments, run as Kaggle competitions

  • Customer exit

    90,000 rows

    8 classifiers

  • Flight ticket price

    40,000 rows

    8 regressors

  • Mushroom edibility

    7,000 rows

    7+ classifiers

Projects on a time axis

  1. IESCP platform May to Aug 2024

  2. Consumer segmentation Oct 2024

  3. Teaching assistant (RAG) Mar to Apr 2025

  4. Comment classification Jan to Mar 2026

  5. Smart MCQ solver May to Jul 2026

  6. VC and HR dashboards Aug 2026

Toolkit, by where I used it. Skills grouped by evidence: used in a project or internship, or covered in a course.

Used in a project or internship on this page Course, certification or exposure

Languages and query

  • Python
  • R
  • DAX
  • TypeScript
  • SQL

Analytics and ML

  • scikit-learn
  • LightGBM
  • XGBoost
  • Optuna
  • Transformers and LoRA
  • Class imbalance
  • OLS regression
  • Chi-square, Kruskal-Wallis
  • K-means

BI

  • Power BI
  • Star schema
  • Advanced Excel
  • Power Query
  • Tableau (course exposure)

Engineering

  • React
  • NestJS
  • FastAPI
  • Flask
  • PostgreSQL
  • SQLite
  • Docker
  • AWS EC2
  • Nginx
  • Vercel
  • GitHub
  • n8n

Business

  • Lean Six Sigma, DMAIC

All projects

Research

Multi-Class Affect Detection in Educational Chatbots: Evidence from Imbalanced Dialogue Data

Accepted, Springer Nature FICTA 2026 proceedings

74.19% recall on the rare Confused class, RoBERTa + SVMSMOTE

Both papers in detail

Next: One assistant became nine

Chapter 5 of 10

One assistant became nine

Today the team is nine agents. Sends, deletes and public pushes wait for my approval.

Runs on open source Hermes Agent; the agent roles, approval gate and routing are my design.

Since 11 Sep 2026, on a cloud server since 20 Sep 2026. Build log

The team: one router and eight agents that own one area each, running on a server
  • Router
  • Academics
  • Career prep
  • Finance
  • Infrastructure
  • Personal ops
  • Socials
  • Consulting
  • 1 more, private
Hours with my agents, by day
peak 13.5 h 2026-09-13 2026-11-01

331.6 hours over 46 active days. Counted from messages I sent, 15 minute idle gap. How these are counted

Which model does the work: share of total logged tokens
  • Gemini Flash 68.4%
  • Claude Opus 18.2%
  • Claude Sonnet 13.1%

Next: What the agents actually do

Chapter 6 of 10

What the agents actually do

A voice note becomes a finished task. Risky steps wait for me.

01 · INGEST Any-device input Dictated on phone or Mac Discord, Telegram 02 · TRIAGE Router agent Reads intent, finds the owner Files a kanban task 03 · DOMAIN AGENTS Specialist agents Always on, on an Azure server Own workspace, memory, tools 04 · MODELS Right model per task Claude or Gemini, chosen by task complexity Plain scripts run before any model WAITS FOR ME 05 · GATE The human gate Pushes, emails, deletions, ledger writes wait for me 06 · DELIVERY Delivered and tracked Google Tasks, GitHub, Firefly A chat note when it is done
Six stages from voice note to delivery
01 Input

Any-Device Ingest

Voice note from phone or Mac

02 Router

Central Triage & Router

Router picks the owner, files a task

03 Agents

Domain agents

Specialists, each with own memory

04 Models

Model choice

Scripts first, stronger models for hard tasks

05 waits for me

The Human Gate

Sends, deletes, pushes wait for me

06 Output

Delivery & Ledger Sync

Results land in Google Tasks, GitHub, the ledger or a chat note.

But can you trust them?

Chapter 7 of 10

Why you can trust the agents

Agents do the work. Layers of checks keep it in check. When a layer misses something, the next one is built to catch it.

Defence in depth means no single control has to be perfect, because another layer is watching. A fail-safe is a backup that acts on its own if one layer misses something.

Five layers between an agent and anything leaving
1 · Gate Needs my approval 2 · Watch Watches for drift 3 · Sweep Checks health hourly 4 · Release Publishes only number updates Your work and output itself 5 · Backup If all else fails: restore

Outer layers check before anything moves. Backup is the way back.

Read the five layers in words
01Gate

Approval gate

Every send, delete and public push waits for my approval.

See the human gate
02Watch

Guardrail detectors

11 detectors watch routing, scheduled jobs, memory, the vault, hub activity and model behaviour.

See the platform
03Sweep

Hourly doctor and cleanup

Health checks every hour. Cleanup is proposed first; I decide.

See the platform
04Release

Protected files and release guard

Instructions change only by proposal. The site publishes automatically only when the change is numbers only.

See the platform
05Backup

Backups

Encrypted, copied offsite, restore tested every week.

See reliability
When a layer fails
1 Gate passed 2 Watch passed 3 Sweep missed 4 Release caught here caught 5 Backup not reached

A miss at one layer is meant to be caught by the next.

What it is built to stop

A leaked key

Control A vault that fills a password only on the matching site

Proof A scan of the public repo found none

An agent doing something unreviewed

Control Approval gate, plus a check for the coordinating agent doing hands-on work

Proof Test faults built in on purpose are caught

A wrong change going live

Control A release check that allows only number updates

Proof 45 of 45 test cases built to fail on purpose behaved as intended

Lost work

Control Encrypted offsite backups, weekly restore check

Proof Backups restored in a weekly test

Settings changing by accident

Control Hourly checks plus daily review

Proof Each hourly and daily run is logged

Five risks, the control that addresses each, and the proof it works.
RiskControlProof
A leaked keyA vault that fills a password only on the matching siteA scan of the public repo found none
An agent doing something unreviewedApproval gate, plus a check for the coordinating agent doing hands-on workTest faults built in on purpose are caught
A wrong change going liveA release check that allows only number updates45 of 45 test cases built to fail on purpose behaved as intended
Lost workEncrypted offsite backups, weekly restore checkBackups restored in a weekly test
Settings changing by accidentHourly checks plus daily reviewEach hourly and daily run is logged
Where the system caught itself

Different layers catch different things, and I publish the misses too.

Layer 4 Release

A refresh that would not publish

A site refresh found it had changed more than numbers, so it refused to publish.

Layer 2 Watch (link check)

A dead link, caught and fixed

A link check found a dead link after release. One commit fixed it.

Layer 2 Watch (text check)

Problems an earlier check missed

A newer diagram-text check found 17 problems an earlier check had missed.

Check it yourself

Reliability Platform Memory

Now see what that team built

Chapter 8 of 10

How the products are built

A closer look at Termly and CogPrep Arcade: the features, the code and the tools behind them.

Two apps anyone can open, and two tools I use

Public beta

Termly

Term planner for IIM Ranchi MBA students.

  • Timetable by section, with free periods marked
  • Absence tracking against the attendance rule
  • Deadline list pulled from the term calendar

Free, ad free, works offline.

Independent student project for IIM Ranchi MBA students. Sign-in: IIM Ranchi accounts. Built with my AI agents.

Open the welcome page

Free to try

CogPrep Arcade

Six aptitude style puzzle games with a coach.

  • A move-by-move coach after every attempt
  • Timed runs, with an optional anonymous score
  • Runs entirely in the browser, no install

Free, in your browser.

An independent practice tool, not affiliated with any employer or assessment provider. Timed runs can post an anonymous score. Built with my AI agents.

Play CogPrep Arcade Source on GitHub

Public code

3 public reposas of 2026-11-01github.com/krshnndu

  • callready

    My repo

    Rehearse a listed company earnings call against its own disclosures: rules-first, AI-assisted, any company

    Last push 2026-10-29 52 commits by me, all time

  • krshnndu

    Profile README

    Writing repo: posts and notes.

    Last push 2026-10-27 7 commits by me, all time

  • aptitude-game-practice

    Source for CogPrep Arcade

    Last push 2026-10-25 36 commits by me, all time

Commits I wrote, last 12 weeks
32411542026-08-102026-09-212026-10-26

92 of 95 commits, all time.

Go deeper

Chapter 9 of 10

Going deeper

Six case studies show the architecture. The code stays private.

The six case studies
Model ladder: first rung that can do the taskSix rungs, tried in order: a script, Gemini 3.8 Flash low, Gemini Flash medium, Claude Sonnet 5.5, Claude Opus 5.5, then Fable at maximum effort for consultation only. FIRST RUNG THAT CAN DO THE TASK 1 A script can do it Script, zero LLM calls 2 Bulk read-only work Gemini 3.8 Flash, low 3 Edit a script can check Gemini Flash, medium 4 Ungated edit, synthesis, software Claude Sonnet 5.5 5 Critical judgment Claude Opus 5.5 6 Novel or irreversible plan Fable, max, consult only
Each task takes the first rung that can do it. See the full routing diagram.

All six case studies

Research on how people use this

Adoption and Implications of Generative AI in Marketing: A Systematic Review and Research Agenda

10th Pan-IIM World Management Conference 2025, IIM Ranchi

Research on how people use this

Affect Detection from Imbalanced Dialogue Data

FICTA 2026, Springer Nature (Scopus-indexed); 74.19% rare-class recall on a 6-class imbalanced dialogue dataset

Next: What I believe and where I am going

Chapter 10 of 10

What I believe and where I am going

Beliefs

Four beliefs, each with proof

  • Agents do the work. I own the decision.

    Sends, deletes and public pushes wait for my approval.

    See the layers

  • Report the honest number.

    Churn model: about F1 0.64 next to accuracy 0.84, because accuracy flatters an imbalanced model.

    See the model

  • Build for the people who use it.

    Termly plans the term; CogPrep Arcade coaches aptitude practice.

    See what I shipped

  • Trust comes from layers, and from publishing the misses.

    Five layers; three misses published in chapter 7.

    See the misses

Direction

Now, until Mar 2027, and after

  1. Now

    Building agent systems and the apps around them, behind a human gate.

  2. Until Mar 2027

    Finishing the MBA at IIM Ranchi, with the placement committee role alongside.

  3. After

    More tools for people who use them.

Read the build log

Who is Krishnendu Biswas?

Krishnendu Biswas is a 2nd-year MBA student at IIM Ranchi (class of 2027) and an AI agent builder who builds multi-agent platforms and runs them daily behind a human approval gate.

He has studied the BS Degree in Data Science and Applications at IIT Madras since May 2023, and completed the Diploma in Programming.

He serves on the Corporate Relations and Placement Committee at IIM Ranchi and is an author of a research paper on dialogue affect detection accepted for publication in Springer Nature FICTA 2026 proceedings, a Scopus indexed series.

Reach me

If this fits what you are building, email me about agent systems, products or research.

Email Resume (PDF)

krishnendu.biswasi22@iimranchi.ac.in

Back to top