Loading

AI engineer. I take systems from schema to screen.

role
AI engineer, Data, AI & Product, Atlys
status
Open to remote roles
time zone
IST, UTC+5:30
education
B.Tech, Mathematics and Computer Science, KIIT, 2027 (expected)

I'm an AI engineer at Atlys, building AI agents on the Data, AI & Product team. Since 2025 I've built multimodal RAG chatbots for The Trav Story, an emergency-vehicle routing system on PostGIS, a fraud-detection ensemble and a marine-species classifier at 98.87% accuracy.

Résumé (opens in a new tab)Email me

I also spent a remote internship at Brndo migrating its site to Next.js and winning back 9.1% of a client's abandoned checkouts. I write Python and TypeScript every day, ship with Docker, and care as much about the API contract and the database schema as about the model.

Diagram: how I structure an AI agentSix parts: input, model, tools, memory, decision and action. Input goes to the model; the model reads context from memory and calls tools; the model's decision triggers an action; the action's result is fed back into memory for the next turn. The highlighted line is the request path: input, model, decision, action. How I structure agents in general; Atlys specifics are not public.callcontextresult fed backinputa message or documentmodelLLMtoolsmemorydecisionaction

Plate 1How I structure agents. Atlys specifics are not public.

  • Agents in production.I build AI agents at Atlys and built the multimodal RAG chatbots behind The Trav Story (LangChain, FastAPI, vector databases).

  • Real-time systems.DIVERT routes first responders around traffic with OSRM, PostGIS and WebSocket tracking at about 500 ms latency.

  • Models that hold up.OceanBot: 98.87% validation accuracy across 31 species. YVOO: 94% with an XGBoost and logistic-regression ensemble.

I build AI that does real work.

  1. AI Engineer, Data, AI & Product

    Atlys

    when
    Present

    I build the AI brains and agents behind Atlys: systems that read documents and conversations, make a decision, and take the next step without a human in the loop.

  2. E-Commerce Intern, SEO & SDE

    Brndo · Ads and marketing

    when
    to
    mode
    remote
    stack
    Next.js, React, Tailwind CSS, Shopify Liquid, Meta Pixel, Google Tag Manager, Zoho CRM, WhatsApp Business flows

    My first fully remote role: two months at Brndo, an ads and marketing agency, working across SEO, web engineering and e-commerce growth for Brndo and its clients.

    • Migrated brndo.in from WordPress to a Next.js single-page app built with Tailwind and React, integrated Zoho CRM's APIs, added server-side rendering and Schema.org markup for SEO, and set up Meta Pixel to track inbound leads through the pipeline.
    • Built Shopify Liquid product modals that give a cricket-equipment brand's buyers a clean, uncluttered product view.
    • Ran predictive analysis on a client's YouTube channel metrics and raised viewer engagement by 12%.
    • Built flows on a WhatsApp marketing platform from website behaviour and user segments, then ran targeted reminder campaigns that recovered 9.1% of abandoned checkouts and drove 21% of an e-commerce client's revenue.

  3. Co-founder & Lead Architect

    The Trav Story · AI-unified travel platform · Bhubaneswar

    when
    to Present
    stack
    LangChain, RAG, FastAPI, Vector databases, Supabase PostgreSQL, Docker, GitHub workflows (CI/CD)

    The Trav Story plans a trip the way a well-travelled friend would: ask a few questions, propose an itinerary, and book it without twelve open tabs. I designed and built the platform.

    • Built multimodal LLM chatbots with LangChain and FastAPI, grounded in our own data through RAG pipelines on vector databases.
    • Deployed the whole stack (frontend, backend, database and the neural models) as distributed cloud services that talk to each other over APIs.
    • Set up the Docker builds and GitHub workflows the team ships through, and designed the Supabase PostgreSQL schemas, with server-side rendering for fast retrieval.
    • Won Best Startup Idea at KIIT University.

  4. Founder & CEO

    DeuxStem · Ed-tech non-profit

    when
    to Present

    I started DeuxStem in 2020 to give students outside big cities the STEM exposure I had to find on my own. It grew to an audience of more than 2 million worldwide.

    • Taught students in more than 10 countries, on every continent.
    • Secured funding and hosted international workshops.
    • Worked with students on improving aviation designs.
    • Led asteroid-search campaigns with the International Astronomical Search Collaboration (NASA-affiliated): more than 20 asteroid detections, VT60 and UF78 among them.
    • Earned commendations from NASA, the Canadian Space Agency (CSA) and the German Aerospace Center (DLR).

  5. Growth Specialist and organizer

    GDG KIIT

    when
    to Present

    I work on growth: getting the right people into the room and keeping them there.

    • Selected in the top 8% of the growth team.
    • Led marketing, outreach and sponsor relations for the Breaking an LLM symposium: 1,500+ registrations.
    • Helped host the chapter's largest hackathons to date, with 600+ attendees; through promotion and sponsorships, grew registrations by 350% and engagement by 84% across events reaching 2,500+ attendees.
    • Competed in Building Bad and joined the Asteroid Hunt deep-tech initiative. Mentored students in cloud and AI and completed 19 Google Cloud courses.

  6. Founding team member

    Startup Grind KIIT

    when
    to Present

    Selected to the founding team building the campus startup community: founder talks.

  7. Project Wing contributor

    Microsoft Learn Student Ambassadors (MLSA)

    when

    Implemented DQN and PPO with Stable-Baselines3 and trained an agent to play Super Mario Bros at an expert level. Taught AI, ML and cloud basics to peers.

  8. AI/ML Researcher

    E-Labs

    when
    to

    Guided student research projects in AI and ML and mentored juniors.

  9. B.Tech in Mathematics and Computer Science

    KIIT University

    when

    Focus: Computational mathematics, distributed backends, ML pipelines, graph and spatial algorithms.

  10. High school, computer science focus

    Delhi Public School

    when
    to

    Focus: Competitive programming and my first real software.

Quiet room, clear problem.

last remote role
Brndo, Jun to Jul 2026
time zone
IST, UTC+5:30

The most productive stretch of my working life so far was two months at Brndo in 2026: fully remote, as an e-commerce intern on SEO and software. I set up somewhere calm and clean, and the time I’d have spent getting to work went into the work itself. No commute, and a schedule shaped around the hours when I think best.

It shows in what shipped. I moved brndo.in from WordPress to a Next.js app with server-side rendering and Schema.org markup, and wired Meta Pixel into the lead pipeline. I built Shopify Liquid product modals for a cricket-equipment brand’s buyers. I ran predictive analysis on a client’s YouTube metrics and raised viewer engagement by 12%. And the WhatsApp reminder flows I built won back 9.1% of abandoned checkouts.

That’s the setup I’m looking for next: a remote AI engineering role on a team that trusts people with outcomes.

A quiet place in Sikkim. My desk photo is coming.

A quiet place in Sikkim. My desk photo is coming.

Time-zone band: my working hours in IST against yours
Show data
Working hours in IST and in your time zone
My time zoneIST, UTC+5:30
My working dayTo be confirmed

Two remote months, measured.

where
Brndo, ads and marketing
mode
remote
length
2 months
when
Jun to Jul 2026

An e-commerce internship in SEO and software engineering, done fully remote for Brndo, an ads and marketing agency, and its clients.

Two remote months at Brndo, measured
ResultWhat it was
WordPress → Next.jsI rebuilt brndo.in as a Next.js single-page app with Tailwind and React, connected to Zoho CRM’s APIs.
SSR + Schema.orgServer-side rendering and structured data, so search engines can read every page.
Meta PixelInbound leads tracked from the first visit through the pipeline.
Shopify LiquidProduct modals that give a cricket-equipment brand’s buyers a clean, uncluttered view.
+12%Viewer engagement on a client’s YouTube channel, after my predictive analysis of its metrics.
9.1%Abandoned checkouts won back by WhatsApp reminder flows I built on a WhatsApp marketing platform from on-site behaviour and user segments.
21%An e-commerce client’s revenue driven by those reminder campaigns.

All of it shipped from a quiet room, on a schedule I set.

What a remote team gets

  1. I own the whole path.Schema, model, API, deployment, interface. Fewer hand-offs means fewer meetings to coordinate them. At Brndo that meant taking brndo.in from WordPress to Next.js, with the SEO and lead tracking built in.

  2. I write it down.Briefs before code, pull requests with context and a test plan, decisions recorded where the team can find them. This site ships with its own architecture document; my LeetCode automation ships with a step-by-step guide.

  3. I manage myself.Solo expeditions are planning exercises with real consequences: route, food, weather windows, turn-back times. I bring the same habits to deadlines.

What I’m looking for

Roles
AI engineer, applied ML engineer, LLM and agent systems, backend for AI products
Location
Remote
Team
Ships to real users, writes things down, measures outcomes rather than hours online.

Hiring remotely?

I’d like to hear what you’re building.

Email meRésumé (opens in a new tab)

Stack

groups
5
in all
52items
every day
7items

Languages

count
5items
  • Python
  • TypeScript
  • SQL
  • C/C++
  • JavaScript

AI and ML

count
11items
  • LangChain
  • AI agents
  • RAG and vector databases
  • PyTorch
  • TensorFlow
  • scikit-learn
  • XGBoost
  • CNNs and LSTMs
  • Transfer learning (InceptionV3)
  • Reinforcement learning (DQN, PPO, Stable-Baselines3)
  • Pandas and EDA

Full stack

count
14items
  • FastAPI
  • Next.js
  • PostgreSQL
  • React
  • Node.js
  • PostGIS
  • Redis
  • Supabase
  • Firebase
  • WebSockets
  • OSRM
  • Tailwind CSS
  • Framer Motion
  • Shopify Liquid

Cloud and DevOps

count
9items
  • Docker
  • GitHub Actions and CI/CD
  • YAML
  • Kafka
  • Jenkins
  • GCP
  • Vercel
  • Render
  • MLOps

Tools

count
13items
  • Git and GitHub
  • Claude Code
  • NotebookLM
  • Google AI Studio
  • Power BI
  • LaTeX
  • Kaggle
  • Google Colab
  • VS Code
  • Figma
  • Meta Pixel
  • Google Tag Manager
  • Zoho CRM

Projects

The Trav Story: An AI travel platform that turns a conversation into a bookable itinerary.

project
The Trav Story
period
Nov 2025 to present
role
Co-founder & Lead Architect
stack
Next.js, FastAPI, LangChain, RAG, Vector databases, PostgreSQL, Supabase, Docker, GitHub Actions, Feature-Sliced Design
links
thetravstory.com (opens in a new tab)Instagram (opens in a new tab)
result
Best Startup Idea, KIIT University

Problem.Planning a trip means hours across blogs, maps and booking sites, and the result still misses what the traveller wanted.

Approach.Multimodal LLM chatbots built with LangChain and FastAPI take the brief, grounded in the platform's own travel data through RAG pipelines on vector databases. The frontend, backend, database and neural models run as distributed cloud services that talk over APIs. Supabase PostgreSQL holds trip state, with server-side rendering for fast retrieval; Docker builds and GitHub workflows ship it.

Result.Best Startup Idea, KIIT University.

Diagram: The Trav Story pipelineA traveller's brief goes to the Next.js app, then to a FastAPI and LangChain service. The service retrieves passages from a vector database of the platform's own travel data and asks the LLM. The itinerary is stored in Supabase PostgreSQL and rendered back to the traveller on the server. Each part runs as its own service, shipped with Docker and GitHub workflows. The highlighted line is the request path.retrievetravellerthe briefNext.jsFastAPILangChainvector DBtravel dataLLMPostgreSQLSupabaseitinerary,server-rendered

System diagram of The Trav Story

Screens to come: The Trav Story, from a chat to an itinerary

DIVERT: Clears a signal corridor for ambulances and fire trucks before they reach the junction.

project
DIVERT
period
Dec 2025 to Apr 2026
stack
FastAPI, Node.js, PostGIS, OSRM, Redis, WebSockets, Docker, Random Forest, XGBoost
result
Working pipeline

Problem.First responders lose critical minutes at red lights and in jams that a routing app can see coming.

Approach.OSRM and PostGIS compute the route and automatically generate diversions of about 800 m around congestion. A Random Forest model predicts the vehicle's ETA at each junction so signals can be pre-empted in time. Dockerised FastAPI and Redis microservices track the vehicle over WebSockets with about 500 ms latency.

Result.Working pipeline built as the Smart City Infrastructure module of my minor project.

Animated diagram: the DIVERT signal corridorA street grid. An emergency vehicle drives east along an avenue; ahead of it, each traffic signal on its route turns to clear in sequence. Where the avenue is jammed, the route takes a diversion of about 800 metres around the congested block, then rejoins the avenue. The final frame shows the whole route drawn and every signal on it cleared. The highlighted line is the vehicle's route.congestiondiversion,about 800 m

System diagram of DIVERT

Accolades: An opportunity-discovery portal: student grants, fellowships and hackathons, searchable in one place.

project
Accolades
period
Dec 2025 to present
role
Founder
stack
Next.js 14, TypeScript, Supabase, PostgreSQL, Framer Motion
links
accolades.site (opens in a new tab)
result
100% data integrity

Problem.The best opportunities for students are scattered across hundreds of sites and usually found a week after the deadline.

Approach.A Next.js 14 portal on Supabase (PostgreSQL) with a navigation-centric routing architecture: server-side rendering for a fast first load, then an interactive Framer Motion "Cyber-Glass" interface rendered on the client. An automated TypeScript ingestion pipeline sanitises and upserts JSON records into a normalised relational schema with strict unique constraints, so a listing appears once and only once. Open Graph metadata, XML sitemaps and schema.org Event JSON-LD make each listing eligible for Google rich results.

Result.Live at accolades.site; 100% data integrity, with zero duplicate listings by construction.

Diagram: Accolades ingestion pipelineJSON records from many sources are sanitised and upserted by a TypeScript pipeline. A unique constraint, drawn as a gate, lets each listing through once and stops duplicates. Listings land in a normalised Supabase PostgreSQL schema, are rendered on the server by Next.js 14, and each listing page carries schema.org Event JSON-LD. The highlighted line is one record's path.jsonjsonjsonrecords frommany sitessanitiseupsertuniqueduplicates stopat the uniqueconstraintPostgreSQLSupabaseNext.js 14server-renderedlistingEvent JSON-LDeligible for rich results

System diagram of Accolades

Screens to come: Accolades search results

OceanBot: Identifies marine species from photos and pulls coordinates out of field reports.

project
OceanBot
period
Jul 2025 to Sep 2025
stack
TensorFlow, InceptionV3, Python, RegEx, Power BI
links
Kaggle notebook (opens in a new tab)
result
98.87% validation accuracy

Problem.Marine survey data arrives as unlabelled images and free-text logs, and both are slow to process by hand.

Approach.Transfer learning on InceptionV3 with a custom head (Dense 1024 ReLU, Dense 31 softmax), L2 regularisation at 0.001 and Dropout at 0.4 to curb overfitting. A language processor with 15+ regular expressions extracts coordinates from text. A Power BI dashboard shows the results.

Result.98.87% validation accuracy across 31 species, 12 points above the baseline. Runner-up (2nd rank) at the PIC conference.

Diagram: OceanBot's classifier headA photo passes through a frozen InceptionV3 base, then a trained head: a Dense layer of 1024 units with ReLU and L2 regularisation of 0.001, Dropout of 0.4, and a Dense layer of 31 units with softmax, one per species. The 31 class cells are drawn evenly. Separately, field reports pass through 15 or more regular expressions that extract coordinates. The highlighted line is the photo's path.Transfer learning: only the head is trained.photoInceptionV3frozenDense 1024ReLU, L2 0.001Dropout 0.4Dense 31softmax31 speciesreportfrom the field15+ regexlat/loncoordinates

System diagram of OceanBot. The class strip is drawn evenly; it shows no per-species accuracy.

Screens to come: the OceanBot dashboard

YVOO: Flags CIBIL-score and income fraud before a loan is approved.

project
YVOO
period
Sep 2025 to Dec 2025
stack
Python, XGBoost, scikit-learn, FastAPI, Docker, Render, React, PostgreSQL
links
GitHub (opens in a new tab)
result
94% and 92% validation accuracy

Problem.Lenders need to spot fabricated credit histories and inflated incomes without real fraud data to learn from.

Approach.I synthesised behavioural datasets, ran EDA to find the signals, and trained an ensemble of XGBoost and logistic regression. Models run behind secured FastAPI endpoints, Dockerised and deployed on Render with auto-scaling.

Result.94% and 92% validation accuracy. 1st rank at the PIC conference. Qualified for Smart India Hackathon 2025.

Diagram: YVOO's fraud-detection ensembleA loan application reaches a secured FastAPI endpoint, which builds its features. The features go to an XGBoost model and a logistic regression in parallel; the two scores are blended into a risk flag that returns to the caller. Both models were trained on synthesised behavioural data. The service is Dockerised and deployed on Render. The highlighted line is the request path.applicationFastAPIsecuredfeaturesXGBoostlogisticregressionblendrisk flagboth trained onsynthesised behavioural data

System diagram of YVOO

Screens to come: the YVOO scoring screen

KANAD: Spots crop disease from a leaf photo and schedules irrigation from the weather.

project
KANAD
stack
CNN, LSTM
result
Winner, NYAS Junior Academy challenge

Problem.Smallholder farmers often notice disease too late and water on habit rather than need.

Approach.A convolutional neural network classifies leaf images by disease; an LSTM forecasts water need from time-series data to schedule irrigation.

Result.Winner, New York Academy of Sciences Junior Academy challenge.

Diagram: KANAD's two models, illustrativeIllustrative. A leaf photo is cut into tiles and a convolutional neural network classifies each tile by disease. Separately, a recurrent LSTM network reads a weather time series and forecasts water need to schedule irrigation. The highlighted line is the leaf photo's path.Two models: one reads the leaf,one reads the weather.leaf phototilesCNNdiseaseper tileweather seriesLSTMrecurrentirrigationwhen to water

System diagram of KANAD. Illustrative. Inputs, crops and dataset are not public yet.

Photo to come: the KANAD prototype

Every build, early experiments included

Honors

  1. New York Academy of Sciences

    New York Academy of Sciences, Junior Academy: Winner

    KANAD, crop-disease detection and irrigation scheduling with CNNs and LSTMs.

    Winner

  2. KIIT University

    Best Startup Idea, KIIT University

    The Trav Story, AI travel platform.

    1st

  3. Smart India Hackathon

    Smart India Hackathon 2025: national qualifier

    YVOO, CIBIL and income fraud detection.

    Qualifier

  4. PIC conference

    PIC conference: 1st and 2nd rank

    YVOO (CIBIL and income fraud detection) took 1st; OceanBot (98.87% species classification) was runner-up.

    1st & 2nd

  5. MIT LaunchX

    MIT LaunchX, 2022: full ride, top 8%

    Awarded a full ride, in the top 8% of applicants. Co-founded Sedu, a game-based health-education startup, with a global cohort; it made $785 in revenue across 4+ countries.

    Full ride

  6. International Astronomical Search Collaboration

    20+ asteroid detections with NASA-affiliated IASC

    Through DeuxStem's search campaigns, VT60 and UF78 among them.

The full register