Work

Agents, vision pipelines, and full-stack products. The numbers are real. Case studies where they exist, source where they do not.

ANU

Apple Swift Student Challenge 2026 winner

A voice-first iOS assistant whose brain runs on the phone: Gemma 4 E4B through LiteRT, on-device classifiers routing every message, and the cloud only as a sanitized, optional fallback.

  • swift 6
  • litert
  • gemma 4 e4b
  • swiftui
Case study
86.7% task classifier, 60-case eval
300+ unit and UI tests green in CI
92.6% routing accuracy
85.3% of queries routed at zero API cost

Dual-LLM Personalization System

Final project for Northeastern’s graduate Generative AI course, written up as a NeurIPS-format paper

A three-layer router that classifies query complexity and dispatches each request to the right model, cutting cascade rate from 30.8% to 14.7%.

  • python
  • scikit-learn
  • gpt-4o
  • streamlit
Case study

Notova

Privacy-first voice notes on iOS and macOS from one shared Swift core, a parallel Kotlin app on Android, and a backend architected so it can never receive audio.

3 platforms, 0 bytes of audio server-side

  • swift 6
  • kotlin
  • typescript
  • oauth2

Meeting Bot

Local-first macOS meeting notetaker. Captures whatever audio is already playing, transcribes on-device with Whisper, ships three front-ends over one engine.

3 front-ends, $0 running cost

  • python
  • whisper
  • swiftui
  • chrome extension

Agent Certifier

Grades an AI agent codebase it did not author across 60 static-analysis tests in 6 weighted categories, without executing a line of it.

60 tests, 10 languages

  • fastapi
  • next.js 14
  • sqlalchemy
  • docker

Agentic Research Assistant

Multi-agent research loop with orchestrator, research, and reasoning agents over dual-layer ChromaDB memory, long-term plus working, retaining context across sessions.

3 agents, ArXiv-grounded

  • python
  • langchain
  • chromadb

AIFORG

A Palantir-style AI platform for a whole company: role-gated agents grounded in its own documents, with 14-category PII tokenization and a full audit trail.

new tool integration under 2 hours

  • python
  • fastapi
  • google adk
  • chromadb

TryOn

Embeddable virtual try-on: a 360 degree phone video becomes a 3D body mesh, a brand’s size chart drives the fit, and a 2D photo mode warps garments onto the shopper’s own photo. Web, Android and iOS SDKs.

video to 3D body mesh, size-accurate fit

  • python
  • fastapi
  • opencv
  • swift
  • kotlin

Real-time Lane Detection

LaneNet encoder-decoder with an EfficientNet backbone, profiled in PyTorch until it ran in real time.

99% accuracy at 150 FPS

  • pytorch
  • lanenet

Vision-Language Grounding

Zero-shot CLIP grounding with cross-modal attention visualization, Stable Diffusion generation, and embedding-space analysis.

4 pipelines, fully zero-shot

  • openclip
  • stable diffusion

AutoFill Pro

Chrome extension that fills job applications across 46 tested platforms from 44 locally stored profile fields, with an optional AI fallback for custom questions. Contact details are redacted on-device before anything is sent to the AI.

46 platforms, 1,300 test cases

  • javascript
  • chrome mv3
  • nvidia nim

SkillGap AI

Skills-gap analysis platform mapping user profiles against live job-market embeddings. Three PMI San Francisco awards as sole developer.

500+ concurrent users

  • react
  • rag
  • postgresql

Real-time Human Pose Estimation

Published at IEEE AISP 2024. A lightweight keypoint regression head on MediaPipe landmark priors, with LSTM activity recognition across 11 activities.

85% mAP, 0.78 IoU, 30+ FPS

  • python
  • mediapipe
  • lstm

RAG from Scratch

Ten commonly recommended RAG techniques implemented and measured on BEIR SciFact against a bge-base dense baseline (nDCG@10 0.759). Seven of them made retrieval worse, one was flat and two helped; the guide walks through each result and why.

10 techniques measured, 8 made it worse

  • python
  • beir
  • bge
  • ndcg@10

Agentic Second Brain

Gemini extraction agent that parses order confirmations into a structured SQLite database, then answers order-history queries and surfaces re-order links through a chat UI.

deployed on DigitalOcean

  • python
  • fastapi
  • gemini
  • next.js

Agent-Guidelines

Guideline-driven multi-agent orchestrator with persistent memory across runs and a live execution dashboard, so agent runs can be watched and diagnosed while they happen.

live execution dashboard

  • python
  • gemini

BugHound

CodePath AI110 capstone. I extended CodePath’s BugHound starter agent with an LLM-output quality gate, import-risk scoring, and a 20-check evaluation harness.

coursework: quality gate + eval harness

  • python
  • gemini

Neural Studio

Infinite-canvas workspace pairing Research and Creator modes on a single surface, built against the Gemini API.

Devpost hackathon build

  • react
  • vite
  • typescript
  • gemini

Hospital Management System

Full-stack clinical portal covering patients, appointments, and billing, with JWT authentication and role-scoped access over MongoDB.

jwt auth, role-scoped access

  • react
  • typescript
  • vite
  • express
  • mongodb

Omni-Axon

Stanford x DeepMind hackathon, April 2026. A master agent classifies each request, then answers, runs a saved workflow, or assembles a team of sub-agents with Drive and Docs tools.

agent builder, reusable workflows

  • typescript
  • react
  • gemini
  • google workspace

Avigis Biopharma

The official website for Avigis Biopharma Private Limited, built and deployed end to end for an external client.

shipped for a client

  • next.js
  • tailwind
  • typescript

More experiments and coursework live on GitHub.

About
Experience, education, and the stack I work in
Contact
Email, LinkedIn, and resume