Hello, I’m
AI engineer · Bank of New York
I build AI products and the backend systems they rely on: agent-based tools, search and retrieval over financial data, and the services underneath. Before that I worked on Photoshop Express for iOS at Adobe and built apps that thousands of students at IIT Guwahati used.
A shared course library for IIT Guwahati
A shared library of lecture slides, assignments, past papers and notes for IIT Guwahati. Students sign in with their institute account, their enrolled courses appear on their own, and every course gets a library they can browse on the web or on their phone.
Deep learning for bubble segmentation in stirred tank reactors
For my B.Tech thesis I benchmarked seven instance segmentation models on high-speed images from a gas-liquid stirred tank, then used their outlines to measure bubble size distributions across operating conditions. Mask2Former was the most accurate and YOLOv8 the fastest, and the comparison has been accepted as a paper in The Canadian Journal of Chemical Engineering.
Bubble size sets how much gas-liquid surface a stirred tank has, and that drives mass transfer. Measuring it from high-speed images is hard because bubbles overlap, deform and break up, so classical methods like the Hough transform or watershed segmentation struggle. The thesis compared seven deep learning models on the same data: Mask R-CNN with and without PointRend, CondInst, SOLOv2, YOLOv8, StarDist and Mask2Former.
The images came from a dual-impeller reactor filmed at several heights, radial positions, stirring speeds and gas flow rates, about 22,500 frames in all, of which 400 were annotated bubble by bubble. Mask2Former reached the highest average precision at 0.651, YOLOv8 was the fastest at 69 ms per image, and models with very different architectures agreed on the bubble size distributions to within about 0.2 mm. The paper has been accepted for publication in The Canadian Journal of Chemical Engineering.
A conversational stylist that knows your wardrobe
Named after the costume designer from The Incredibles, Edna is a fashion assistant you can chat with. Tell it where you are going and it puts together complete looks from your preferences, past purchases and current trends, then reworks them as you react. It also keeps a virtual wardrobe built from photos or a few words, suggests an outfit of the day from clothes you already own, and points out what your wardrobe is missing.
Every recommendation starts in a prompt generator that blends what Edna knows about the user, such as age, region, fit, budget and favourite colours and brands, with fashion trends of the moment. GPT-3.5 answers with a few looks in a fixed structure, a parser splits each look into items, and every item is matched to real products with a price, rating and photo, so you can pick the version you like best.
The wardrobe fills itself from photos. A U2-Net model fine-tuned on the iMaterialist dataset separates each piece of clothing, and BLIP-2 captions it with its colour, pattern and type. Trends come from a scraper that reads fashion blogs, Instagram and Pinterest, with a fine-tuned DistilBERT model summarising them into tags the prompt generator can use. The app is written in Flutter and talks to a Django server backed by SQLite. I built it in a team of three.
Routing riders and sizing parcels for last-mile delivery
Grow Simplee’s problem at the Inter IIT Tech Meet: plan the day’s routes for a fleet of delivery riders and measure every parcel before it leaves the hub. Our team modelled routing as a vehicle routing problem with pickups and deliveries in OR-Tools, estimated parcel volume from a depth camera over a mock conveyor, and tied it together in a manager dashboard and a rider app, much of which I built.
Routing is a capacitated vehicle routing problem with mixed pickups and deliveries, solved with OR-Tools on distance and time matrices from OSRM. Pickups and deliveries carry opposite weights so a rider’s bag never overflows, and routes are capped in length and duration. When the solver has to leave an order for tomorrow, a drop penalty decides which one: it grows as the delivery date gets closer and with how busy the order’s neighbourhood is, measured on Bangalore’s ward map. New pickups arrive during the day, so routes are rebuilt every 15 minutes from the riders’ live positions.
A sweep-based clustering step keeps large days fast, cutting route planning for 2,000 orders from 300 seconds to 60. On standard benchmarks the solver landed within 4 to 12% of the best known solutions. For parcel volume, a RealSense depth camera, U2-Net segmentation and a small shape classifier handled boxes, cylinders, spheres and cones with about 2% error, against 18% from point clouds. I worked mostly on the rider app and the manager dashboard in React and Redux: the full mobile delivery flow, rider management, the live maps, bag packing views and their integration with the Django backend.
Stamp duty, worked out through a few questions
Stampez works out the stamp duty on a legal document. The amount depends on the state, the type of instrument and details that change from case to case, so a Python pipeline reads each state’s stamp act and uses Gemini to turn every instrument into a branching questionnaire. You pick a document, answer a few questions and get the duty along with the article of the act it comes from.
Stamp acts are long schedules of instruments, each with its own conditions, rates and references to other articles. The pipeline pulls a state’s schedule out of the official PDF and HTML tables, splits it into instruments and converts each one into structured JSON. It then maps the cross-references between instruments as a dependency graph, asks Gemini to write a questionnaire for each instrument, merges the referenced rules back in and optimises the flow so nobody answers more questions than they need to.
The finished questionnaires live in MongoDB. A Next.js admin app lets the team review and correct them state by state, and the public site, also built in Next.js, walks you through the questions and shows the calculation and its legal reference at the end. Everything ships in Docker containers.
Your courses
assets/projects/edna-mode.png
+3 More projects
Machine learning
Handwritten equations to LaTeX
An encoder-decoder model with attention, written in PyTorch, that reads a photo of a handwritten equation and writes out its LaTeX. OpenCV cleans up the image first, and a small Flask app puts the model behind an upload page.
Android
The official app for IIT Guwahati’s cultural festival
Event schedules you could personalise, e-passes, a merchandise store and plenty of custom animation, built natively for Android. It passed 3,000 downloads with a 4.9 star rating during the festival.
Android and web
Mock JEE tests from scanned question sheets
An Android app scans a sheet of printed questions, splits it into separate questions and converts them to LaTeX with OCR. A companion website then serves them in a copy of the real JEE exam interface, with a question palette, sections and instant checking.
Make it work, make it right, make it fast.
02 Experience
Current
Jul 2024 to present · Pune
Bank of New York
I build AI-powered wealth management products and the services behind them, used by financial advisors and their clients.
May to Jul 2023 · Noida
Adobe
I worked on Photoshop Express for iOS with the UX, engineering and product teams.
03 Background
I started programming in 8th grade, building small Android apps, and slowly worked my way out to websites and backends. When I got to IIT Guwahati to study Chemical Engineering, with a minor in Mathematics, most of my time still went into software.
A lot of that time went into the Coding Club, where I ended up heading app development. I taught the First Steps to Flutter course to more than 2,500 students and led the teams that built CourseHub and AsterHF. I also led the official Alcheringa app, which passed 3,000 downloads. In 2023 our team won gold at the Inter-IIT Tech Meet, and I won Best Project Overall (Software) at Techevince the same year. My B.Tech thesis turned into a computer vision study of bubbles in stirred tank reactors, which has since been accepted by The Canadian Journal of Chemical Engineering.
Since 2024 I have worked as an AI engineer, mostly on agent systems, retrieval and the services around them. Outside work I read about new AI techniques, help a few developers who are starting out, and keep some side projects going.
04 Toolkit
Hover over a toolTap a tool to see where I have used it, and the field will redraw its mark.
Pick a tool
Python, Java, Flutter…
05 Contact
Email is the quickest way to reach me. My code is on GitHub and my work history is on LinkedIn.