Medium ↗6 min read
How Diamond answers (AI ChatBot)
Diamond answers from a small set of approved Bhutan Airlines Thailand pages and drops citations that were not in that set.
Ideas in chronological order
Essays
Technical essays, research notes, and field reports published here and on Medium.
The useful assistant is the one that can stop
Diamond answers Bhutan Airlines visitors from a bounded packet of site pages, drops citations that were not retrieved, and hands off when the pages do not support the claim.
Medium ↗6 min read
The useful AI job agent is the one that refuses to apply
A job-search agent that can click Submit is easy to demo. CareerBot scores fit, preserves evidence, and refuses to apply for you.
CareerBot: prepare auditable job applications without auto-apply
CareerBot is an open-source MCP toolkit that scores jobs against a Master CV, drafts checked resumes, and saves SHA-256 snapshots. It does not scrape boards or submit applications.
Medium ↗12 min read
Token Economics for LLM Agents: A Practical Guide for Applied AI Engineers
A practical guide to token budgets, memory, observability, tools, security, and multi-agent system economics.
Dopamine.Diet: A Local-First Productivity System for Serial Project Switchers
A local-first deep-work dashboard for people who keep switching projects before anything ships.
Parameter Golf: From Zero to a Working Cloud Baseline
Reproducing OpenAI's Parameter Golf language model training challenge on RunPod — covering environment setup, SSH debugging, dataset wiring, and landing a val_bpb baseline on a single H100.
Reproducing the LPCVC 2026 Track 1 Baseline on Qualcomm AI Hub
A walkthrough of reproducing a working CLIP image-to-text retrieval baseline on XR2 Gen 2, covering environment fixes, SSL certificates, dataset paths, and preprocessing gotchas.
I Built 9 AI Agents to Write My Thesis — Here's How It Works
A first-time thesis writer, not a strong coder, built an open-source AI agent system that helped him publish in a Q1 journal and defend his thesis. This is the full story.
A Framework for Flying Smarter: The Practical Deployment Guide
After 8 GSD levels, 6 models, and 3,368 tile-label pairs, here's the condensed operational guide for deploying AI-based oil palm monitoring with a drone.
From Pixels to Meters: The Math Behind Canopy Size Estimation
A segmentation mask is a collection of pixels. Here's how we convert that into a physical canopy area in square metres and an equivalent diameter — with the full derivation and error analysis.
The Hybrid Paradox: Why SAM + YOLO Sounds Better Than It Is
SAM 2.1 was trained on 1 billion masks. My 56MB YOLO model still outperformed the hybrid pipeline in most conditions. Here's the detailed explanation.
Six Models Enter, One Problem Wins: A Head-to-Head Segmentation Showdown
I trained YOLOv8, YOLOv11, Mask R-CNN, and two SAM hybrid pipelines on oil palm drone data. Here's what the numbers actually showed — including the result that surprised me most.
Building the Dataset Nobody Had: Annotation, Corruption, and the Split That Kept Everything Honest
How I built a multi-resolution oil palm segmentation dataset from scratch — including the 301 corrupted label files I had to fix before anything could run.
The Drone That Couldn't See: Why Altitude Breaks Your AI Model
A model trained at 3cm resolution failed at 6cm. Same drone. Same plantation. Different altitude. Here's the physics and math behind why — and what to do about it.
I Built an Obsidian Plugin to Publish to Jekyll
An open-source Obsidian plugin that converts wiki-links, image embeds, and callouts to Jekyll-compatible Markdown — then copies everything to your _posts/ folder.
Why YouTube Comments Are Disabled on University Wi‑Fi
How content filters cause it and why a VPN fixes it.
How I Built This Blog in One Session — With AI as My Co-Pilot
From zero to a fully-featured minimalist blog using Jekyll, GitHub Pages, and AI pair programming. A transparent look at ethical AI-assisted development.