AI | Rabimba Karanjai

Racecraft: Fine-Tuning Gemma 4 with DPO for Sub-150ms On-Device Coaching at 130 MPH

Sun, 02 Aug 2026 00:00:00 +0000

Racecraft: Tale of Car Sensors

Sat, 01 Aug 2026 00:00:00 +0000

AEGIS-CHAOS: From 'Vibe Coding' to Closed-Loop SRE

Sun, 19 Jul 2026 00:00:00 +0000

Splitting the Brain to Beat the Clock

Thu, 25 Jun 2026 00:00:00 +0000

Teaching the Coach to Read the Driver

Mon, 15 Jun 2026 00:00:00 +0000

The 800-Millisecond Problem

Wed, 10 Jun 2026 00:00:00 +0000

Google I/O GDE Field Test — AI Race Coach at Sonoma Raceway

Sat, 23 May 2026 08:00:00 +0000

Dr.MRI.AI

Mon, 18 May 2026 00:00:00 +0000

Dr.MRI.AI is a web-based DICOM viewer built on a simple observation: the hard part of medical AI isn’t the model, it’s knowing what to send it. A typical MRI study contains 200+ slices, but only a handful carry the signal. Dr.MRI.AI plans the slice selection before any multimodal analysis runs.

How it works — Plan → Review → Synthesize:

Plan: Gemma 4 produces a structured plan (which series, slice range, window/level) from your clinical question in plain English. You can accept, edit, or reject it.
Review: The multimodal model analyzes only the focused subset — 200+ slices reduced to ~16 frames, ~99% fewer image tokens.
Synthesize: A final text-only pass produces a structured report with clickable slice references; export findings plus reviewed evidence as one bundle.

Privacy-first architecture:

DICOM files are processed entirely in the browser — no data is uploaded to any server
Fully browser-local inference via onnx-community/gemma-4-E2B-it-ONNX through Transformers.js + WebGPU
Provider-agnostic: Gemini API, OpenAI-compatible endpoints, Ollama (MedGemma planning + Gemma 4 review), or fully on-device Gemma 4 Browser

Built with React 19 + TypeScript + Vite and Cornerstone3D for medical image rendering. Featured in the Google Developer Experts May–Jun 2026 AI community roundup. Educational/research tool — not certified as a medical device.

My Friend's MRI Didn't Come with a Manual, So I Built One with AI

Mon, 18 May 2026 00:00:00 +0000

AI Race Coach (Project Koru)

Fri, 01 May 2026 00:00:00 +0000

The AI Race Coach (Project Koru) is a real-time coaching system for motorsports, developed during the Google I/O GDE Field Test at Sonoma Raceway. The project demonstrated how agentic AI can bridge domain gaps between software engineering and racing expertise.

Architecture:

Antigravity — stateful orchestration and telemetry ingestion
Gemini API — cloud-based driver modeling and strategic insights
Gemma 4 12B — on-device edge intelligence for zero-latency audio coaching
Jetpack Compose — high-refresh Android dashboard for real-time visualization
Google ADK — multi-agent orchestration framework
Custom USB Interface — 10 Hz telemetry stream wired directly from vehicle sensors to Pixel 10

The system identified performance gains beyond human perception — including a new throttle application zone mid-corner in Turn 2 that secured a 0.1-second advantage. Featured on the Google Developers Blog and covered in my blog post Racecraft: The Origin Story.

Rethinking Publication: A Certification Framework for AI-Enabled Research

Thu, 23 Apr 2026 00:00:00 +0000

CausalGraphX: A Counterfactual Graph Neural Network Framework for Explainable Systemic Risk Assessment

Sun, 01 Feb 2026 00:00:00 +0000

Building Production AI Systems with Free Cloud GPUs

Tue, 18 Nov 2025 00:00:00 +0000

Build Smarter AI Agents Faster: Introducing the Google Agent Development Kit (ADK)

Wed, 09 Apr 2025 00:00:00 +0000

Deep Dive into the Google Agent Development Kit (ADK): Features and Code Examples

Wed, 09 Apr 2025 00:00:00 +0000

Artificial Intelligence-Based Target for Personalized Interventions of Atherosclerosis from Gut Microbiota Signature

Sat, 01 Feb 2025 00:00:00 +0000

LLM Reasoning & Code Generation

Wed, 01 Jan 2025 00:00:00 +0000

My research on LLMs spans reasoning evaluation, automated code/test generation, and reliability. Key contributions include:

AgentTester — A multi-agent framework for automated unit test generation that significantly improves software reliability (AIWare 2025).
HPCAgentTester — Extending multi-agent test generation to high-performance computing environments.
LogBabylon — A unified framework for cross-log file integration and analysis using LLMs.
Hallucination Mitigation — Techniques for reducing AI hallucinations in diagnostic and clinical contexts.
Evaluating LLM Rationality & Randomness — Benchmarking the quality of reasoning and entropy in LLM-supported tasks.
Synthesizing Public Opinions — Using LLMs to synthesize democratic discourse (IEEE ICEDEG 2025).

This line of research explores how multi-agent systems can produce more reliable, verifiable AI outputs, with applications from software engineering to healthcare.

Quantum Computing & AI

Wed, 01 Jan 2025 00:00:00 +0000

Exploring the intersection of quantum computing and machine learning:

QuCoWE — Quantum Contrastive Embeddings with Variational Circuits for near-term quantum devices (AAAI 2026). This work introduces a novel approach to learning embeddings using quantum variational circuits, designed to work within the constraints of current NISQ-era hardware.
Texas Quantum Initiative — Serving on the Scientific Advisory Committee, helping shape the state’s quantum research strategy.

This emerging research direction investigates how quantum advantages can be practically realized for machine learning tasks, even with today’s noisy intermediate-scale quantum devices.

Enhancing Vascular Disease Diagnosis through AI-Driven Analysis of Histopathology Images

Wed, 01 May 2024 00:00:00 +0000

Mitigating Hallucinations in AI-Driven Medical Diagnosis

Wed, 01 May 2024 00:00:00 +0000