Skip to content

Software Engineer @ Kim CC

Devroop
Saha

AI Engineer

I build AI systems that actually make it to production: retrieval that retrieves, agents that finish the job, and latency numbers I'm willing to put on a slide.

See the workGet in touchRésumé

01 / about

Engineer by trade,
culé by temperament

The work

I'm an AI engineer who likes the unglamorous half of the job: the retrieval that returns the right chunk, the guardrail that catches the bad reply before a customer sees it, the Docker image that went from 1GB to 158MB because someone actually looked.

Most of my work lives in production LLM systems: RAG pipelines, multi-step agents, graph-based workflow engines, hybrid search. I'm a software engineer at Kim CC in Bengaluru, where I joined as an intern and took the AI support platform from handling 18% of workflows to 60%, then pushed product-search accuracy from 20% to 100% with dense and sparse retrieval on Qdrant.

These days the work is broader than models: persona and response-strategy training that lets non-engineers set the brand voice, an eval cascade that scores every layer of the pipeline, guardrails and confidence gates deciding when the AI should keep quiet, and multi-brand support for accounts running several brands at once. Before Kim CC there was an AI assessment tool at DailyWellnessAI (and a 90s to 30s latency cut that I enjoyed far too much), freelance RAG work on GST documents, and a generative flashcard pipeline. I also write about all of it on Medium, because explaining a thing is how I find out whether I understood it.

The rest of it

Off the clock, I am almost entirely a football problem.

Lionel Messi is the GOAT. This is not a debate I'm hosting, it's a fact I'm announcing. I've watched grown men argue about trophies and nations and eras and none of it survives thirty seconds of him picking up the ball on the right touchline and deciding the game is over now. Barça is home: the good years, the very bad years, the ones where I had no business staying up until 1:30am for a league game and did it anyway.

If I'm not writing code, the odds are extremely high that I'm watching football. Not even good football, necessarily. A mid-table game on a wet Tuesday will do. My weekends are scheduled around kick-off times and I've made peace with it.

Outside of that: The Dark Knight is my favourite film and I'll rewatch it at the smallest provocation. It's the rare blockbuster that respects you enough to be about something. I grind LeetCode (300+ and counting, though the counter matters less than the fact I stopped being scared of DP). And I write, mostly to force my own thinking into a straight line.

LeetCode solved
300+LeetCode solved
Medium articles
8Medium articles
Hackathon finishes
3Hackathon finishes
Public repos
59Public repos

02 / experience

Where I've shipped

Four companies, one theme: taking LLM systems from a demo that impresses to a service that holds up on a Tuesday afternoon under real traffic.

  1. Kim CC

    Jul 2026 – Present

    Software Engineer · Bengaluru, India

    • Built Persona training and intent-scoped response strategies end to end: LLM generation pipelines, CRUD APIs, and full create/dashboard/validation UIs, consumed by the Python AI engine, so non-engineers control brand voice and how the AI answers each intent.
    • Built the AI Operating Settings surface: per-intent filtering toggles over a ported intent classifier, all 9 system guardrails configurable per account with policy assembled from the database, auto thresholds, and a confidence check and stale-response gate deciding when the AI answers and when it escalates.
    • Delivered multi-brand support across the platform for enterprise accounts running several brands in one workspace: brand management with inbox assignment, server-to-server normalizer APIs resolving each incoming ticket to a brand, and brand-scoped analytics and agent scorecards.
    • Designed a 6-layer eval cascade for the AI engine with LLM-as-judge scoring and Langfuse telemetry, turning response quality into per-layer metrics that pinpoint which stage produced a bad answer.
    • Instrumented the engine feature by feature with metrics across guardrails, knowledge base, rule-engine workers and tool runners, and defined the alert rules on top, so each feature alarms on its own failure modes instead of one blanket service alarm.
    • Surfaced the AI's thinking steps as live progress events from the Python engine through to the agent UI, and added visitor page-trail tracking to live chat so agents can see which pages a customer browsed before opening a ticket.
    • Integrated Skio, Trustpilot, Klaviyo, Yotpo, Reviews.io, USPS, ShipStation, ShipBob and Google Sheets into the agent tool layer.
    • Improved the Pulse analytics product with sentiment and per-account tag filters, background CSV exports, and LLM-analysed CX reports emailed as PDF.

  2. Kim CC

    Nov 2025 – Jun 2026

    AI Engineering Intern · Bengaluru, India

    • Lifted AI response accuracy and workflow coverage from 18% to 60% by redesigning extraction logic, retrieval workflows, validation layers, and fallback handling.
    • Implemented hybrid product search with Qdrant dense and sparse retrieval for Shopify product discovery, taking retrieval accuracy from 20% to 100%.
    • Developed a graph-based workflow execution engine orchestrating multi-step AI pipelines, tool calls, routing logic, and async task execution.
    • Shipped webhook ingestion pipelines for Zoho and Gorgias plus a Zoho Desk extension, letting agents generate, review, and send AI-assisted replies inside the helpdesk UI.

  3. DailyWellnessAI

    Mar 2025 – Aug 2025

    AI Engineering Intern · Remote (San Diego, CA)

    • Developed an AI assessment tool using the OpenAI SDK with Structured Output to evaluate free-form Q&A and recommend tailored service packages.
    • Cut agent processing latency from 90s to 30s by introducing threading for concurrent tool calls and batching where possible.
    • Orchestrated deployment with FastAPI, Docker, and AWS EC2, compressing images from 1GB to 158MB and establishing CI/CD.

  4. GST Magic AI

    Aug 2025 – Sep 2025

    Freelance AI Engineer · Remote

    • Engineered a document management tool for GST PDFs with embedded metadata, supporting search, edit, and status tracking, persisted to Pinecone.
    • Deployed a RAG-based assistant on the Pinecone Assistant API and improved retrieval relevance through metadata augmentation.

  5. Mermory

    Oct 2024 – Nov 2024

    Freelance AI Engineer · Remote (U.S.)

    • Executed a generative flashcard pipeline using LLaMA 3.1 and FastAPI to create cloze-deletion and Q&A flashcards.
    • Designed an occlusion-based approach with OpenCV and NumPy to auto-generate labelled diagram study aids.
    • Created a styled image generation tool with diffusion models fine-tuned via DreamBooth, alongside GANs.

  6. Ajay Kumar Garg Engineering College

    Nov 2022 – Jun 2026

    B.Tech, Computer Science · Ghaziabad, Uttar Pradesh

    8.55 CGPA · 2nd rank in department

03 / resume

The one-pager

Everything above, compressed onto a single side of A4 for the people who need it that way.

Download PDFOpen in new tab
Devroop_Saha_Resume.pdfTap to open · 1 page · 100 KB

04 / selected work

Things I built

Six of them. Voice agents, MCP servers, a CLI people actually pip install, and some computer vision that judges your squat form.

Julius AI

1

An AI interviewer that actually conducts the interview. A voice agent runs a 6-stage conversational screen, then hands you a curated coding challenge, then evaluates both and makes a hiring recommendation. The hard part was latency: 800ms end-to-end over WebSockets with Deepgram STT and ElevenLabs TTS, because anything slower stops feeling like a conversation.

  • Next.js
  • Node.js
  • MongoDB
  • Redis
  • Deepgram
  • ElevenLabs
SourceWatch demoTypeScript

TexMCP

11

A FastMCP microservice that renders LaTeX to PDF, exposed as MCP tools so any model-context-protocol client, Claude Desktop included, can typeset documents on demand. Small, boring, and it does exactly one thing properly.

  • Python
  • FastMCP
  • LaTeX
  • MCP
SourcePython

fastapi-scaffold

10

A CLI that generates a FastAPI project the way you'd actually structure one: optional auth, database wiring, ML model setup, Docker. Published on PyPI and installed by people I've never met, which is still the most satisfying metric I have.

  • Python
  • FastAPI
  • CLI
  • Docker
  • PyPI
SourceView on PyPIPython

YogaFix

3

Real-time yoga pose detection and correction. Frames are captured server-side, run through MediaPipe pose estimation, and feedback streams back over WebSockets fast enough to correct you mid-pose rather than after it.

  • Python
  • OpenCV
  • MediaPipe
  • FastAPI
  • WebSockets
SourceWatch demoPython

FitVid

2

An AI-powered visual gym: it watches your exercise form, counts reps, and tells you when your last three were rubbish. Computer vision aimed at the specific problem of training alone with nobody to correct you.

  • Python
  • Computer Vision
  • MediaPipe
  • OpenCV
SourceRead the write-upJupyter Notebook

Elevate

1

Disk scheduling algorithms (FCFS, SSTF, SCAN, LOOK and their circular variants) explained through the analogy of elevators in a building, and animated with Manim. Built because reading the pseudocode never made it click, and watching it move did.

  • Python
  • Manim
  • Operating Systems
  • Algorithms
SourceRead the write-upJupyter Notebook
All 59 repositories on GitHub

05 / writing

I write things down

Explaining a concept is how I find out whether I actually understood it. Longer experiment write-ups live here on the site; the rest are pulled live from my Medium feed.

The LayaChess board in a browser: a game position next to panels showing Laya's win chance for its candidate moves and Stockfish's preferred move.

LayaChess: teaching a System 1 decision model to play chess

An experiment: I took Laya, a fast decision model that had never seen a chessboard, trained it on 2 million Stockfish-rated moves, and wrapped it in a search engine. Here's how it works, and how strong it got.

  • Experiment
  • Chess
  • Decision models
Read the write-up

How I Cut AI Agent Latency from 90s to 20s Using Python Threads

How I transformed a sluggish AI agent into a performance powerhouse using Python’s parallelization capabilities Picture this: you’re building an AI agent that needs to make dozens…

Read on Medium

From Disks to Elevators: Applying Scheduling Algorithms for Optimal Movement

Elevators are such an excellent invention of human civilization. Imagine with a press of a button, you are at the topmost floor of 50 story building in a matter of minutes. One tim…

  • operating-systems
  • algorithms
  • low-level-programming
Read on Medium

Leveraging Language Models for Advanced Mathematical Computation and Problem Solving

In 2024, with the rising popularity of Generative AI and AI being a buzzword, LLMs are everywhere. And the reason LLMs are everywhere is because they are like a ‘Jack of all trades…

  • mathematics
  • llm
  • ai
Read on Medium

Enhancing Code Documentation with RAG-Based Q&A Systems: A New Frontier in Developer Support

So as developers, when we are developing software, we all have to take the help of the documentation, it’s a cannon event. Even though ChatGPT is there to help developers, there is…

  • python
  • llm-applications
  • generative-ai-use-cases
Read on Medium

FitVid: AI-powered visual gym

This article explains my latest project: FitVid, an AI-powered visual gym that tracks the movements of the gym exercises, counts the repetitions, and helps improve the form of gym…

Read on Medium

A Comprehensive Guide to CNNs Part-1: The Fundamentals

So, What is CNN? In fact, what is a Neural Network for that matter? Well, a Neural network is like an artificial brain that can intelligently do stuff. Computer Scientists have cre…

  • neural-networks
  • deep-learning
  • artificial-intelligence
Read on Medium

Detailed Explanation, In-depth Intuition, and Python implementation of Bagging and Random Forest

In this article, we will dive into the concept of Bagging and study one of the most famous and used algorithms in Industry and Kaggle competitions: Random Forests. The only pre-req…

  • machine-learning
  • ai
  • mathematics
Read on Medium

Understanding Categorical Encoding

In this article, we will try to understand what is a categorical data, what is categorical Encoding, why to perform encoding and How to perform encoding. I will try to explain all…

  • data-preprocessing
  • statistics
  • data-science
Read on Medium
Everything on Medium

06 / toolkit

What I work with

Languages

  • Python
  • TypeScript
  • JavaScript
  • Go
  • SQL
  • HTML
  • CSS

AI / ML

  • PyTorch
  • TensorFlow
  • scikit-learn
  • LangChain
  • LangGraph
  • RAG
  • AI Agents
  • NLP
  • Fine-tuning

Frameworks

  • FastAPI
  • Next.js
  • Node.js
  • Flask
  • Django
  • Streamlit
  • Pandas

Data & Infra

  • PostgreSQL
  • MongoDB
  • Redis
  • Qdrant
  • Pinecone
  • FAISS
  • Docker
  • AWS
  • Git

07 / receipts

Things I won

  • Runner-up

    Sparrowthon (Techdome)

    All-India hackathon, 350+ participants, with Chirpy (RAG for API management).

  • 4th

    Data Analytics Competition, NSSC, IIT Kharagpur

    Out of 450+ participants.

  • Top 10

    AI Bioinnovate Hackathon (IIT Jodhpur & ChemBioAI)

    Against 500+ competitors, on molecular toxicity prediction from SMILES notation.

  • 2nd

    AKTU Technical, Literary & Management Fest

    KIET Ghaziabad, 100+ participants.

  • 2nd

    Department rank in university examinations

    Plus cash awards for academic performance and attendance.

08 / contact

Say something

Hiring, collaborating, or just want to argue about the GOAT debate. The form goes straight to my inbox.