M.Tech CSE, NIT CalicutChennai, India
Engineering resilient software and robust AI for production at scale.
Tailor this page for
M.Tech student at NIT Calicut. Before that, a year and a half at BNY putting LLM tooling and automation into production, and four years of undergraduate research: 12 conference papers, 4 indexed book chapters and a design patent.
Where my code met real users: production AI at BNY, and earlier work at Cognizant and Stanford.
Developer Analyst, Production Services (AI/MLOps, automation monitoring), Chennai
Analyst Intern, UK Asset Servicing Engineering, Chennai
CAFÉ sponsored project, team lead of 4
Section Leader, CS106A
Each one is judged by a number that's hard to fake. Most of them run live on this page.
Distributed key-value store
A 5-node cluster that agrees on every write using Raft consensus: leader election, log replication and snapshotting, written from scratch. Stress-tested by crashing nodes and partitioning the network mid-traffic, then checking every history for linearizability.
Payments platform
A Stripe-style double-entry ledger with idempotent transfers and outbox events, so money is never created or lost on retries. Webhooks are signed and retried with a dead-letter queue, so receivers that go down still get every event.
Try to break it: retry a payment, fire 20 at once, or take the merchant's server down
| Account | Balance |
|---|
Every payment writes balanced debit and credit lines in one transaction, keyed by an idempotency key, so a retried request returns the original result instead of charging twice. Events leave through an outbox and are retried with exponential backoff; after 5 failures they park in a dead-letter queue until replayed. Time is sped up.
Low-latency matching engine
A price-time priority matching engine in modern C++ with no allocations on the hot path and lock-free queues between stages. Fuzz-tested for correctness and replayed against 10 million real Binance L2 events.
Send a market order and watch it sweep the levels
A simulated feed of limit orders, cancels and aggressive orders, matched with real price-time priority: best price first, then first in line at that price. The C++ engine does this 10 million times a second; this one runs about 8 times a second so you can follow it.
Suicide-risk detection with a support chatbot
An attention-based transformer that flags suicide risk in social-media text, paired with a chatbot that responds with support. Tuned for recall on the minority class, because a missed at-risk post costs far more than a false alarm. Accepted at IEEE ETIS 2025.
Drag the threshold and watch missed posts trade off against false alarms
Illustrative score distributions for 1,000 posts, 12% of them at risk; these are not SafeMind's real outputs. They show why the model was tuned for minority-class recall (91% in testing) rather than raw accuracy, which would look high even if every at-risk post were missed.
Real-time vision and language on a phone
A mobile app that looks at a scene and describes it aloud. GAN-based enhancement cleans up dark or noisy frames, a detector finds objects, and a ResNet-LSTM captioner turns them into a sentence, trained on a custom-labelled dataset. Selected for presentation at AICCC 2024 in Tokyo; a journal version (Retina) adds transformers.
Step through what the app does between the camera and your ears
An illustrated frame stands in for the camera. The stages, their order and the kinds of outputs match the real app; the scores shown are examples.
Multimodal fusion for brain image segmentation
A CNN that fuses MRI, CT and PET scans before segmenting, because each modality sees something the others miss: MRI shows soft tissue, CT shows bone, and PET shows metabolic activity. Fusing them gives clinicians one clearer view of a lesion.
Change how much each scan contributes and watch the lesion boundary appear
Synthetic scans drawn in your browser, not patient data. The segmentation is a confidence threshold on the fused image, so it only captures the whole lesion when the modalities agree.
Power-usage forecasting, Centre of Excellence in Data Science
A memory-enhanced radial basis function network that forecasts electricity use for college buildings, so loads can be planned rather than reacted to. The metacognitive part decides what to learn, what to skip, and when to add a neuron, which is where the training-time savings come from.
Switch models and hover the chart to compare forecast with actual
Actual loadForecast
Synthetic campus load, not the original dataset. The gap between the two models is illustrative of the reported improvement.
Smaller explorations that shaped how I think about language, accessibility and trust in AI.
Edge AI and IoT obstacle sensing for visually impaired users. Granted a design patent by the Patent Office, Government of India.
Combines multilingual text, audio and visual cues to read sarcasm, hesitation and empathy in low-resource settings.
Builds medical knowledge graphs automatically with named-entity recognition and relation extraction.
Transformer models plus human-in-the-loop feedback to cut hallucinations and bias in generative AI answers. A paper on it is in progress.
Flags sexist and racist slurs in multilingual corporate text, with feedback that adapts sensitivity across cultures.
Student portfolio, recruitment and placement-prediction tool built with industry support from QuickPlay. Best Project Presentation, 2023.
* Highlighted to match the view you picked at the top.
Type help in the terminal to explore it your way.
Undergraduate work in applied deep learning, mostly for healthcare and accessibility.
Twelve conference papers across IEEE, Springer and Elsevier venues, twelve book chapters (four indexed, eight in review or indexing), one design patent and two journal papers in progress. One paper won Best Paper in the AI/ML track at ICCCNet 2023 in Manchester. Google Scholar profile
PublishedAccepted or presentedIn progressPatent
ICCCNet 2023, Manchester Metropolitan University. Springer LNNS
Published
I-SMAC 2024, Nepal. IEEE
ICSSES 2024. IEEE
ICCCI 2023. IEEE
ICCCI 2023. IEEE
ETIS 2025, Thiruvananthapuram. IEEE
ICSSAS 2024. IEEE
ICCCNet 2024. Elsevier SSRN
ICCCI 2024. IEEE
ICCCI 2024. IEEE
ICRTC 2024. Springer
Journal paper, extending my final-year thesis
Paper in preparation, from the HumaineAI work
Design patent, The Patent Office, Government of India. Edge AI and IoT
CRC Press, Artificial Intelligence for Precision Agriculture
CRC Press, Artificial Intelligence for Precision Agriculture
Scrivener Publishing
Scrivener Publishing
Scrivener Publishing
Cambridge University Press
Cambridge University Press
De Gruyter
Wiley
IGI Global
Nothing in this category yet. Choose All to see everything.
Awards from industry juries, conferences and my college.
The teaching, community work and interests that a one-page resume leaves out.
Teaching is how I check I've actually understood something. Since 2020 I've run sections, workshops and clubs alongside coursework and work.
Section Leader, 2021
One of 1,100 volunteers chosen worldwide to teach CS106A in Python. Led weekly sections from Karel the robot through control flow, decomposition, lists and dictionaries, and classes, and debugged students' code one-on-one.
Co-founder
Grew the chapter to 600+ members with hands-on technical workshops. It has since split into data science, machine learning and communications tracks.
Founding member and mentor
Mentored three student teams whose builds placed in the top 100 of GDSC hackathon challenges.
Founding president and mentor
Ran hands-on workshops on generative AI, LLMs, machine learning and deep learning, from training a first classifier to prompting and evaluating LLMs.
Founding member
Brought together 50 students who wanted to publish, and helped start the open-source club. With IntellexAI, GDSC and IEEE CS, these groups now have about 1,250 members.
Automation mentor
Mentored the RPA team building UiPath bots, and ran the club's events and outreach.
Developer
Rebuilt the college website, the cultural-fest pages and the admin system: attendance, on-duty requests, results, course registration and placements. Two of my papers came out of this work.
Session lead, 2024 to 2025
Taught colleagues AI and ML fundamentals, generative AI and LLMs, deep learning, probability and statistics for ML, and privacy, ethics and responsible AI.
Also: CSE branch prefect on the student council from 2020 to 2024, and a mentor for first-year CSE students.
My oldest habit. Most Carnatic compositions sit on Adi talam, an 8-beat cycle you keep with your hand: a clap and three finger counts, then clap-wave twice.
Trinity College London Grade 6 in plectrum guitar. Western rhythm after years of tala keeps both kinds of timing honest.
I watch films for how they're built as much as for the story, and I'll argue about an ending longer than the film ran.
Questions about mind, ethics and what it means to understand something. They became practical once I started working on AI.
Consumed in roughly equal measure. Hackathons, coding contests and late builds have been a constant since my first year.
Best Student Award for leadership and academics at St. John's Senior Secondary School and The Hindu School, long before any club presidencies.
Curious by default, always mid-build, and powered by code and caffeine in equal measure.
M.Tech, Computer Science and Engineering
Semester 1 courses, with credits
B.E., Computer Science and Engineering
I'm in campus placements at NIT Calicut this cycle for SDE and AI/ML roles. I'm also glad to talk about research collaborations, hackathon teams, open-source work, or a long systems-design conversation.