What five weeks of local coding agents taught me
Speculative decoding, thinking budgets, KV cache settings and a harness that does most of the work. Lessons from running coding agents on my own Mac, my pi config, and where local still falls short.
Writing
Notes on software, AI, open technology, and the questions I can’t leave alone. Real projects, decisions, mistakes, and things learned along the way.
The year the agents got out
My talk at SNEVic's 2026 AGM: how far AI agents have come, what I learned building a Nepali news app, and why checking the work matters more than ever.
Six local AI models on my M1 Max: llama.cpp vs MLX
Six open-weights models on a 32 GB M1 Max. Power settings, speculative decoding and what the benchmark numbers leave out.
Claude Fable 5 lived for 3 days. I spent them upgrading my harness.
June 2026: Anthropic shipped Claude Fable 5, then the US government switched it off three days later. Here's what I did with a frontier model on a countdown, and the lesson I kept long after it vanished.
The time a Guardian reporter thought I went to war with OpenAI
A weekend RAG project I shipped in 2023 came back to almost make me the protagonist of a Guardian story about going to war with OpenAI. It wasn't me, but the actual case is more interesting than the headline.
How I picked a Nepali voice for kchakhabar.com (and cut the bill by 90%)
K cha khabar ships AI-narrated short videos, in two languages. The first week I ran them on ElevenLabs TTS. Then I rebuilt the voice for one-tenth the cost.
The cheapest good enough model is the one that ships. K cha khabar now runs on Gemini 3
A few weeks ago I shipped kchakhabar.com, a cross-publisher Nepali news intelligence platform. Here is the technical story of what's under the hood, and why production runs on Gemini Flash instead of the model that topped the benchmark.
Foundation Models Explained: From AI's Origins to Talking to an LLM (Free Video Series)
A four-part video series I built in 2025 as part of my Agentic AI Software Engineer curriculum, covering AI history, how foundation models are trained, post-training alignment, and hands-on interaction with LLMs.
That mountain on the Narayanghat horizon is Manaslu, the world's 8th highest peak
I grew up in Chitwan watching a peak hover above the northern hills every clear winter morning. I finally worked out what it was.
Agentic DevOps - I gave my Agent root access to a VM. Here is what I learned.
What I learned using AI agents to troubleshoot Kubernetes and manage a virtual machine over SSH, including how I decide what access to give them.