// projects

Things I
Actually Built

Seven projects. Each one has a demo or results panel below — not marketing copy, just what the project actually does.

01 / 07
AI Infrastructure · Multi-Agent Systems

NoteMind FEATURED

Most multi-agent AI frameworks hit a wall: each agent works in an isolated context window, repeating everything to every other agent until you hit token limits. NoteMind replaces that with a persistent semantic knowledge graph.

Each agent writes atomic Markdown notes stored in .brain/, embedded into ChromaDB, indexed in MongoDB, and immediately searchable by every other agent in the swarm. Instead of receiving a 40k-token dump of prior history, an agent gets the 3 most semantically relevant notes. Token cost stays O(k) regardless of swarm size.

Next.js 14FastAPI ChromaDBMongoDB React FlowWebSockets Gemini 2.5TypeScript
Live Simulation — Agent Swarm Graph

TOKEN COST COMPARISON
Standard (5 agents)O(n²) — hits context limit
NoteMind (5 agents)O(k) — stays bounded
02 / 07
Computer Vision · Chrome Extension · Gemini Vision

MangaLens SHIPPED v1.0.8

Google Translate can translate text on a webpage — but it skips over images entirely. That's useless for manga, where the story is inside speech bubbles drawn on the image. MangaLens fixes that.

The Chrome extension detects text regions in manga images using Gemini Vision, fills speech bubbles with their estimated background colour, and overlays translated text in-place. Supports 18 languages, handles lazy-loaded images, builds a translation memory so character names stay consistent across pages.

JavaScriptChrome Extension MV3 Gemini VisionCanvas API Service WorkerCI/CD
Demo — Before / After Translation
Gemini Vision detects bounding boxes → fills bubble bg → overlays translation in-place on canvas

Works with MangaDex, Webtoon, MangaPlus
Handles lazy-loaded / infinite scroll pages
~ Rate limited: 15 images/min on Gemini free tier
03 / 07
Machine Learning · Finance · Reliability

MarketLens

A CNN + Transformer ensemble for stock price forecasting — but the actual point is the RAPID diagnostic framework built on top of it. RAPID scores Reliability, Accuracy, Precision, Interpretation, and Distribution to tell you objectively when the model's output should be trusted.

The output is three confidence bands (P10/P50/P90) rather than a single prediction. The question the project asks: "should we even trust this forecast?" — before you act on it.

PyTorchTransformers CNNRAPID Framework Probabilistic Forecasting
Scenario Simulator — Select Market Condition

RAPID TRUST SCORE
Reliability
Accuracy
Precision
Select a scenario →
04 / 07
NLP · Sentence Embeddings · Streamlit

AI Resume
Analyzer

Takes a resume and a job description, outputs a match score plus which skills are present, which are missing, and why. Uses a three-signal hybrid: TF-IDF (keyword matching), sentence embeddings via all-MiniLM-L6-v2 (semantic similarity), and Jaccard skill overlap.

Built with a Streamlit interface and hardware-aware OpenVINO inference (falls back to CPU automatically). Tested on 2,400 resumes and a LinkedIn job postings dataset.

sentence-transformersTF-IDF Scikit-learnOpenVINO StreamlitPython
Try It — Paste Your Skills
YOUR SKILLS (comma separated)
JOB REQUIRES
06 / 07
Computer Vision · Full Stack · Deployed

TeachAI LIVE ↗

Smart attendance system combining face recognition with GPS geofencing. A student is marked present only when two conditions are met simultaneously: their face matches the enrolled records, and their device is within the classroom's GPS radius. No manual roll-call. No proxy attendance.

Face RecognitionTypeScript GPS GeofencingVercel
Attendance Pipeline Simulation
📸
Camera ready. Waiting for student scan.
07 / 07
Research · Cryptography · Published

Ternary LFSR
Cipher

Published research in Advances in Decision Sciences (Q1 Scopus, CiteScore 8.3). Proposes a GF(3) — base-3 — Linear Feedback Shift Register for image encryption. Unlike binary LFSRs, the ternary feedback mechanism triples the keyspace per stage, making brute-force exponentially harder.

Validated across NPCR (98.04%), Shannon entropy (6.80 bits), UACI, and pixel correlation. Primary application: protecting sensitive medical images during transmission.

GF(3) FieldsLFSR Stream CipherNPCR 98.04% Q1 Scopus
GF(3) Register — Step Through Live
8-stage register state (values: 0, 1, 2)

OUTPUT STREAM (last 24)

PIXEL ENCRYPTION PREVIEW