A tweet, a scanned tax form, a hand sign — trained into signal, shipped as working software. Data Scientist & ML Engineer based in Kozhikode, Kerala.
I'm a computer science engineer who kept saying yes to whatever layer of the stack needed attention — training the model, building the API around it, keeping the server it runs on alive.
Before any of the projects on this page: a year keeping a 100+ camera government AI-surveillance network at 99.5% uptime, then three months at Ladder7 shipping four sentiment products end-to-end, scraper to model to dashboard. One of my earlier builds, a real-time sign-language translator, was published at an IEEE conference and won Best Paper at my university's project expo.
I'm based in Kozhikode, Kerala, and I've just wrapped up at Ladder7 — looking for what's next in software engineering, backend, or ML. I'm not fussy about the title. I want to work on something real.
Thirteen shipped repositories. Four are below in full — the rest are grouped by what they have in common.
A 3-tier document-AI system that pulls structured data — GSTIN, invoice values, tax splits — out of scanned GST 2A statements and E-Way Bills, using a triple-OCR pipeline (PaddleOCR + Tesseract + EasyOCR) feeding a fine-tuned LayoutLMv3 model. Shadow correction and auto-straightening handle poor scans and phone photos; an auto-retraining loop learns from every user correction; output lands as a ready-to-use Excel report.
A unit-normalized analytics dashboard that turns platform trust & safety transparency reports (Meta, X, WhatsApp) into anomaly detection with Isolation Forest, per-organization forecasting with 95% confidence bands, and category-level breakdowns. The data-cleaning layer keeps absolute counts and percentage/rate metrics from ever being wrongly summed together, and an offline rule-based Q&A assistant answers questions directly from the filtered dataset.
Point it at a company website and it auto-discovers the product line, then aggregates customer reviews across five platforms — Google, Twitter, Instagram, YouTube, Reddit — in parallel. RoBERTa-based sentiment and aspect-level analysis runs on the combined review set, and the tool generates a downloadable PDF report summarizing brand- and product-level sentiment.
A real-time ASL-to-text/speech system built around a custom CNN trained on 31,000+ hand-gesture images, with MediaPipe handling hand tracking. Published at IEEE ICCCI 2023 and awarded Best Paper at the KARE Project Expo the same year.
Four independent apps built on the same pattern — scrape a platform, score it with a Twitter-tuned RoBERTa model, lay it out on a dashboard — each adapted to its source's particular mess.
Random Forest regression (R² 0.93, MAE ₹0.45L) estimating used-car resale prices from 7 features, served through a Flask REST API.
RepoA Netflix-style browsing experience built from scratch — auth, watchlist, seeded catalog, live search, Top 10 row.
RepoDesktop Java Swing app: generate, salt-hash encrypt, store, and retrieve passwords, fully offline. Presented at ICCES 2022.
RepoA cyberpunk-themed calculator — real-time computation, full keyboard support, glowing animated UI, deployed on Render.
RepoScrapes any webpage with Playwright and generates an AI summary through LangChain + Groq, in one command.
RepoSoftware engineering, backend, or ML. I'm not fussy about the title — I want to work on something real.
arjunkalliyadath2001@gmail.com