AI/ML Portfolio · Applied Systems
End-to-end ML pipelines, deployed AI systems, and cross-platform products — from transformer theory to production inference.
Technical Stack
Every project is engineered end-to-end — mobile apps, browser automation, multi-agent platforms, and theoretical ML research — with a preference for self-healing, robust systems that handle model output unpredictability gracefully.
Stack: PyTorch · HuggingFace Transformers · scikit-learn · NumPy for ML, Cloudflare Workers · FastAPI · Flask · Supabase for infrastructure, Kotlin · Flutter · Electron · Playwright for cross-platform delivery.
Full monorepo: Stremini AI 15-agent workspace, AI keyboard (StreminiIME), Electron desktop automation agent, Stremini Academy (517 topics, 25 tracks, 6 tiers), and KhetiAI crop disease detection with APMC mandi price forecasting.
DistilBERT IT-support fine-tuning (86.7% acc.), character-level LSTM in pure NumPy, modular CNN on CIFAR-10 (87.3% acc.), 4-layer neural net from scratch, and a Prompt Engineering Lab with automated LLM-as-judge scoring pipeline.
ML Scam Detector pipeline (94.96% accuracy, 28k URLs), interactive HTML papers on Gradient Descent as a Dynamical System and Spectral Properties of Attention Matrices, plus a FastAPI inference endpoint for the deployed model.
Three full-stack AI products: STYFI fashion AI backend (Gemini 2.5 Flash + two-stage SDXL enhancement), Scholira scholarship RAG engine (110 scholarships, TF-IDF from scratch), and Zz-AI desktop voice agent with tiered LLM dispatch.
Restaurant management SaaS with React + TypeScript + Vite + Supabase. RLS policy consolidation, multi-tenant auth, post-signin null restaurant state resolution, and production-ready role-based access control.
Live portfolio housing interactive papers, case studies, project writeups, and a trust score demo deployed as a live Cloudflare Worker endpoint.
Production AI ecosystem: 15-agent web workspace (Intelligence, Creation, Strategy, Execution layers), Android AI keyboard (StreminiIME), desktop Electron automation agent with observe-act loop, and Stremini Academy — 517 topics across 25 tracks and 6 tiers.
AI-powered crop intelligence for underserved Indian farmers. Dual-model architecture — Gemini 2.5 Flash for vision, K2-Think-v2 for reasoning — on a React 19 + Cloudflare Worker stack.
End-to-end interpretable scam detection: 16 hand-crafted features with Logistic Regression + Platt calibration, and a malicious URL detector trained on 28,000 URLs (URLhaus + Tranco Top 1M). Deployed as live Cloudflare Worker.
Three full-stack AI products. STYFI: fashion AI backend with Gemini 2.5 Flash + two-stage SDXL enhancement. Scholira: scholarship retrieval with TF-IDF from scratch (110 scholarships). Zz-AI: voice-first desktop agent with tiered LLM dispatch.
Transfer learning on a 300-sample IT support dataset for 5-class ticket classification. Fine-tuned DistilBERT with a custom classification head, published to HuggingFace Hub.
Custom dataset generation, transfer learning on the classification head, and a lightweight inference API for raw text to ticket category prediction.
View in repo →1.2M parameter CNN with batch normalisation, OneCycleLR scheduler, and label smoothing. Fully modularised: data, model, train, evaluate, visualize modules.
Full LSTM in pure NumPy: all gates, BPTT, gradient clipping, and Adam. Temperature-scaled text generation. Weights serialised as .npz with loss history tracking.
Forward pass, backpropagation, mini-batch SGD (batch=32, 500 epochs). All matrix ops explicit: activations, cross-entropy loss, live loss/accuracy curves.
8 controlled A/B experiments isolating few-shot, CoT, persona, format constraints, and task decomposition vs zero-shot baseline. Automated LLM-as-judge pipeline with 4-axis rubric on a 1-10 scale using K2-Think-v2. Exportable as JSON/Markdown.
Get in Touch
Open to ML research collaborations, AI safety discussions, internship opportunities, and conversations about building technology for underserved communities.
Context & Background
Models in Production
MBZUAI K2-Think-v2 · Gemini 2.5 Flash · Groq Llama 4 Scout · DistilBERT · Groq Llama 3.3-70b
Research Interests
Spectral theory of neural networks · AI safety and prompt injection defence · Interpretable ML for security · Multi-agent architectures · Agricultural AI for South Asia
Collaboration Areas
AI for social impact in South Asia · Educational AI platforms · AI safety tooling · Applied ML research with deployable outcomes
Languages
Python · JavaScript / TypeScript · Kotlin · Dart (Flutter)