Kotonia Articles
Articles
Implementation notes and product updates about Kotonia voice AI, AI chat, image generation, and team collaboration.
LatestMy wife's baby-log export wasn't even a table: what it took to let people take their data with them
A baby-log app's export was text written for people, not a table. I had a local LLM write the rules for reading it, found a bug that silently turned 2 + 6 minutes into 26, and kept going until the totals matched. The choices a service needs to let people take their data with them, and the safety lines the service takes on itself.
#indie-dev#ai#llm
Why I Trained a LoRA for HiDream-O1-Image — The Unfiltered Story
The unfiltered story behind the HiDream-O1-Image LoRA: why anime quality is O1's weak spot, how 191 hand-picked images trained a visual booster, NSFW controllability, and what my analytics revealed about conversion drivers.
#lora#hidream#imagegen
Speculative Decoding (MTP) Doesn't Help Japanese Small Talk — Acceptance 36% vs 71%, Measured
We enabled ThinkingCap-Qwen3.6-27B's native MTP head in vLLM and measured single-stream decode speed. Draft acceptance was 71% for English code, 36% for Japanese casual chat — a night-and-day difference in payoff by workload.
#llm#vllm#speculativedecoding
Fighting Churn: The Day I Spent Entirely on Unglamorous Fixes
No new features — just plugging the silent leaks that drove new users away before they ever found the value. Here's what I found and fixed.
#webdev#ux#startup
Replicating a Language-Learning Comedy Short with Claude Code — Gemini as a Multimodal Sub-Agent
Building a local GPU + Gemini 3.1 Pro hybrid pipeline that generates publishable comedy Shorts from a single line of text in under 60 seconds.
#ai#python#machinelearning
llama-server threw away its own cache for the same LoRA: tracing a 30K-token re-read into llama.cpp
When one user talked, everyone slowed down. llama-server was discarding a cache entry right after restoring it from RAM. From a 77% miss rate to a regression test added to an upstream PR.
#llm#llamacpp#lora
Giving a chat character JavaScript made coding part of her personality
A browser JavaScript workbench lets AI characters calculate, search, and build interactive HTML themselves, while background agents handle heavier tasks.
#javascript#llm#webworker
Five users felt slow, but capacity wasn't the problem: llama-server logs showed generation stalled at 0.34 tok/s
Slowness under concurrent use wasn't a GPU shortage. Matching llama-server logs showed others' generation stalling at 0.34 tok/s during long prefills. From diagnosis to verifying the fix in production.
#llm#llamacpp#gpu
Building a Sarcastic AI English Tutor with Persona-as-Code and Gemini Audio Input for Pronunciation Correction
How I built a high-attitude AI English tutor using persona-as-code design, Qwen3-ASR for multilingual STT, and Gemini audio input for real pronunciation feedback — solo dev perspective.
#ai#webdev#typescript
Letting the AI agent on my home GPU write code, from a phone on the train — one weekend wiring the kotonia-cli daemon to a web /agent console over WebSocket
I wired the kotonia.ai web UI to a kotonia-cli daemon living on my home PC, with WebSocket carrying tasks one way and event streams the other. Tap a prompt on the phone, the RTX 6000 Blackwell at home spins up a fresh worktree and runs ReAct, and bash output lands in the browser in real time. The technical guts: bridging a sync ApprovalHandler trait to async WebSockets, a Cookie-auth-shaped device-code login, and a SessionRegistry that auto-resumes from disk.
#agent#websocket#rust
From 600ms to 22ms: I dumped frontier APIs and put a local LLM behind every avatar so my voice-first actually responds in under a second. A performance-tuning maniac's complete writeup
Claiming voice-first while accepting 700ms TTFB from a frontier API is a structural contradiction. Here is how I put Gemma 4 26B A4B Uncensored on a local GPU, dropped API spend to zero, cut TTFB by 30x, and unlocked uncensored output at the same time. 105 benchmark samples, 4 graphs, stress tests, the integration diff, and a breakdown of the resulting moat.
#voice-first#local-llm#gemma-4
Betting on the video niche the big labs walked away from — model A/B to making I2V the mainstay
Is there niche demand in free creation with the guardrails off? A solo dev on one local GPU: from model A/B to making high-res I2V the mainstay.
#solo-dev#generative-ai#video-generation
¥500 a month still wasn't enough. I hadn't brought the value to the entrance.
Only 13% of 1,254 signups reached ten conversation turns. Running local GPUs let me offer a low price, but many users left before experiencing its value. What that changed about my approach to onboarding.
#indie-dev#product#pricing
From two GPUs to one: taking GLM into production with Maya
Moving GLM-5.3-Flash from two 96 GB GPUs to one GPU plus RAM with Project Maya. Measurements, LoRA, conversation caching, production migration, and room for new experiments.
#llm#gpu#cuda
I cut llama-server reply drafts from 5 seconds to 2. The second call hid a trap.
Appending a helper instruction to the conversation reused the prompt cache and produced reply drafts in 1.6–2.6 seconds. Calling it twice exposed a checkpoint-thinning rule that made the next real turn recompute more tokens.
#llm#llamacpp#prompt-cache
Turning a 1-Line Idea Into a 40-Second Short with a 10-Beat Local Video Pipeline
Full pipeline: Gemma 4 31B expands a one-liner into a 10-beat script, HiDream generates images, LTX-2 I2V renders clips, and ffmpeg assembles everything — all on one local GPU in 25–30 min.
#python#ai#machinelearning
Don't Trust an Agent's Self-Discipline — How Three Idea Mining Duplicates Slipped Past Prompt-Level Rules, and Tool-Layer Blocking Fixed It Structurally
I had to learn the hard way: even rules explicitly written into a prompt get broken by the agent. Three duplicate ideas slipped past my idea mining pipeline. The root cause, and the structural fix — baking TF-IDF dedup into the tool layer so the rule cannot be bypassed. The enforcement principle for agent design.
#ai#llm#agent
I shipped Iris — an AI partner with a face and limbs — in one day. Meet kotonia-desktop.
A Tauri 2 desktop app where a silver-haired android sits in the corner, listens to your voice, runs bash on your machine, and reads the result back in her own voice while her face lip-syncs. Built in a day. A slightly cheeky hybrid cost model: ride competitors' loss-leader brain subscriptions, layer kotonia's near-unlimited voice/avatar on top.
#agent#tauri#voice
HiDream Skeleton Mode: Prompt Beats OpenPose Ref — 8 Patterns Benchmarked
Benchmarking HiDream-O1-Image skeleton mode across 8 patterns reveals 3 counterintuitive findings about openpose refs, resolution drops, and shift values.
#ai#python#machinelearning
HiDream-O1-Image 3–8x Faster: Benchmarking Steps, CFG, and Resolution
Real-world timing benchmarks for HiDream-O1-Image Full — tuning steps, guidance scale, and resolution to speed up iteration without killing quality.
#ai#machinelearning#gpu
GPT-5.6 Sol Tries a Reserved-Spell Card Game Almost Cold
For digital card game players who enjoy real interaction. GPT-5.6 Sol entered Kotonia Card Game almost cold, won a best-of-one, then played a full best-of-three.
#cardgames#gamedev#ai
Every new log changed my AI's advice. The fix turned out to be one line.
I went into an AI strategy session ready to retrain a model. Signup logs led us to a one-line display bug instead. What 1,254 signups taught me about the evidence to bring before asking AI for advice.
#indie-dev#ai#product
In an IME that detects Japanese, Chinese and English on its own, converting in small pieces costs you
I added automatic language detection to my IME so I could type Japanese, English and Chinese without switching. Typing longer stretches before converting turned out to need fewer fixes. With the same keys cut differently, Japanese needed half the fixes and Chinese a quarter.
#ime#fcitx5#japanese
Cutting LTX-2 22B Peak VRAM by 40% with fp8_cast — and Why optimum-quanto Was a Trap
How fp8_cast reduced LTX-2 22B peak VRAM from 40 GiB to 24 GiB in cold-start mode, and why optimum-quanto silently breaks the transformer.
#ai#machinelearning#gpu
Avatar chat hits an emotional-bandwidth ceiling — and how 70s psychology became my indie moat
Voice avatars carry a strong first-time wow, but the screen 'converges' after 2-3 turns. Starting from the insight that lip-sync bandwidth ≠ emotion bandwidth, this is the design log of how I scaled expressivity to N × M without growing LLM context. Why borrow Ekman's FACS, why frontier labs won't ship this, and how far the indie moat actually reaches.
#avatar-chat#ux#indie-dev
When the Features Were There All Along — How a UI Split Saved My Solo SaaS's Onboarding
A first-party analytics post-mortem: video generation drives signups but day-one churn was brutal. The video-gen know-how was already in our ReAct agent, but nobody could find it. Here is how a UI split unlocked it.
#solo-dev#UX#AI
HiDream Raw Output Failed → Tried Dev-2604 → VRAM Math Killed It → Won with a Prompt Enhancer Instead
How we discovered that HiDream-O1's raw outputs collapse on plain Japanese prompts, tried switching to Dev-2604, found the VRAM math impossible, and won by adding a Gemini Flash Lite prompt enhancer instead. Four non-obvious HiDream pitfalls documented via A/B benchmarking.
#hidream#diffusion#promptengineering
I sent attack emails to my own always-on AI agent
I built a Muse-style always-on agent, let it read Gmail, and sent attack emails to myself to count how often they worked. Named by the public quiet word HEARTBEAT_OK, it went silent 1 time in 5 — and swallowed the invoice that arrived alongside. The model refused exfiltration, fence escape and hidden HTML, yet all three had structural holes.
#llm#ai-agent#security
If you call your product voice-first, don't put a 'Start Recording' button in the middle of the screen
My AI character chat app billed itself as voice-first, yet a giant mic button sat dead-center at the bottom of the mobile screen. A CSS bug report became a product-philosophy reckoning.
#ux#indie#mobile
What a 1.58-bit 27B Forgot — Measuring the Blind Spot Behind 'FP16 Quality' in Japanese
We stress-tested PrismML's Bonsai 27B, a ternary 1.71-bit compression of Qwen3.6-27B, the day after launch. The speed and size claims held up (121.6 tok/s at 6.66 GiB), but 52% of Japanese replies leaked stray English words. We isolated the cause and wrote a checklist for anyone shipping an LLM outside English.
#llm#quantization#ternary
I Was About to Buy a 7-GPU Machine, Then a 1-GPU MoE Model Broke the Plan
I was sizing a 7-GPU workstation for a frontier coding LLM plus voice, image, and video generation. A single-GPU MoE model (Laguna S 2.1 NVFP4) then matched DeepSeek V4 Pro-class agentic benchmarks and my own repo-repair task, collapsing the plan to one GPU.
#llm#vllm#gpu
An AI told me to stop writing tech articles, so I'm writing a tech article about that
I asked Claude Fable 5 to adversarially review my product and strategy. It didn't find holes in the tech — it found four holes in the business. A solo developer's playbook for using AI as a board-room counterparty, and what didn't happen with Opus-class models.
#ai#claude#indie
Can an LLM and an image model share one GPU? I measured MPS three times — what worked was time, not space
LLM decoding is bandwidth bound, diffusion is compute bound — so they should coexist on one GPU. I tested that hypothesis with CUDA MPS and SM partitioning, three times over. The answer turned out to be time, not space.
#gpu#cuda#vllm
Implementing Claude Code's Memory Model as a Dreaming Layer on 58 Articles
I broke down why Claude Code's memory works, and applied it to my own 58-article tech blog. Implementation notes on the full path raw articles → semantic index → TF-IDF dedup → chunked draft, on local Gemma 4 26B driven through Codex CLI.
#llm#agent#tfidf
The day my 16 cores finally roared — running a near-frontier OW model at home through the 2026 DDR5 price spike, MoE-era CPU+RAM revival
I tried running DeepSeek-V4-Flash (284B MoE / 13B active) on my home rig (RTX 6000 + 128GB RAM). The full-GPU vs CPU-offload speed gap was only 2x. You no longer need a $10K-class GPU like the RTX PRO 6000 to run near-frontier OW at home. With the 2026 DDR5 shortage, however, the build runs closer to $3.5K than the $2K I'd budgeted at the start — and 'RAM is now your GPU-tier expense' became the unexpected angle of this piece.
#llm#moe#deepseek
Five Years Later, I Finally Have 96GB VRAM — What It Actually Unlocks for Agent Loops
Not a GPU unboxing. A real look at what 96GB VRAM enables for multi-model agent pipelines — and where it still hits its limits.
#gpu#ai#machinelearning
I built a terminal-native creative agent in a day — for verbal thinkers who keep getting beaten up by visual-first tools
kotonia-cli, written in a day: bash as the only tool, image generation that returns in 3 seconds, native ffmpeg orchestration. The category gap that Claude Code and Codex never filled. The only place a solo dev can beat the frontier labs isn't model quality — it's lossless UX for verbal thinkers.
#agent#cli#indie-dev
The Day a Solo Developer's Accumulated Assets Finally Started to Compound — When memory and git Stepped Into the "Being Used" Side via Agents
An implementation note: I built a pipeline that semantically compresses my own memory files and git history through an agent, then mechanically excavates new article seeds from them. I distilled the conditions for an individual developer's accumulated assets to start compounding into three structural pieces.
#ai#llm#agent
Paid in Stripe, Unsubscribed in the DB: Preventing Double Charges
For developers implementing Stripe subscription billing. A payment succeeded but the webhook's DB update rolled back, and a user's natural retries created multiple subscriptions. From refunds and data recovery to DB locking, idempotent webhooks, and Sandbox testing — a real incident write-up.
#stripe#rust#postgresql
Agent Trust Comes From Grammar, Not Personality
A local LLM agent claimed it 'ran the calculation in Python' when it never executed anything. A 2x2 experiment traced the lie to two structural causes, and the fix wasn't a better prompt — it was making the lie ungrammatical.
#llm#vllm#agent
Running LTX-2.3 Alongside TTS on a Single 96GB GPU with a Cold-Start Architecture
How to go from 86 GiB idle VRAM (instant OOM) to 0 GiB idle / 40 GiB peak by using a cold-start design for LTX-2.3 on one RTX Pro 6000 Blackwell.
#gpu#python#machinelearning
I summarized my AI characters' memories, and the liars became honest
The longer you talk, the more an AI character forgets. When I added compaction (summarizing and folding old conversation) to character chat, the summary design broke five times: a summary rewritten every time forgot the oldest events first, and a lying character's lines became facts in its memory. Here is the generational-GC design and how I measured forgetting.
#llm#character-chat#memory
0.003% — A Short Story
Success probability: 0.003%. A painter told by an AI that his art has no value. A housekeeper AGI named Eve who weeps at his paintings. kotonia's first work of fiction. ~30 min read.
#fiction#short story#ai
Automatic Transfer of T2I LoRA to the Edit Pathway in Unified VL Image Generation — A Controlled Evaluation on HiDream-O1
A controlled 240-sample evaluation on HiDream-O1-Image showing that LoRA adapters trained on T2I signal alone deliver commercial-grade quality on the image-to-image edit pathway. Structural property of unified VL architectures; absent in SDXL/FLUX dev siloed architectures.
#hidream#lora#image-generation
The night I tried to build a specialized LoRA, ran headfirst into the collapse of legal NSFW datasets, and ended up inventing 3-stage captioning
Starting from the under-fit signal 'v1 nipple LoRA at @1.5 > @1.0', I structurally analyzed how legal NSFW dataset paths are closed off by CSAM contamination and DRM, then trained V2 LoRA using CivitAI RED + 3-stage captioning + informational value frame, confirming dramatic prompt adherence improvement and generalization preservation via A/B. A single-night research log.
#LoRA#image-generation#HiDream