ControlMT documentation — the journey + the playbook
A first-person record of building a 139M-parameter Kannada↔English translator from scratch, solo, on one consumer GPU, with an AI assistant as a collaborator. For ML engineers who want to build something similar.
This folder is six self-contained docs. Each one stands alone — read whichever match your interest, in any order. They cross-reference each other so you don't have to read all of them to learn one thing.
Pick what you want to read
| If you want to … | Read this |
|---|---|
| Get the highest-density version of everything I learned | top-lessons.md — 10 lessons, one paragraph each |
| Read the full chronological story (v1 → v2.3 public release) | the-journey.md — what I built, when, why, what surprised me |
| Skip my mistakes — what I tried that didn't work | what-didnt-work.md — 8 failed experiments + root-cause analysis |
| Just the concrete recipes (no theory) | how-it-was-built.md — data filtering, training schedule, eval, deployment |
| Learn the AI-assistant collaboration patterns | working-with-claude.md — memory rules, background tasks, what to delegate |
| Find anything in the repo | repo-map.md — folder layout + file conventions |
Reading paths by use case
"I have 10 minutes."
Read top-lessons.md. It's the synthesis. Each lesson links to the doc that explains it in depth.
"I want to build a small specialized translation model."
how-it-was-built.md— concrete pipelinewhat-didnt-work.md— skip my dead endstop-lessons.md— the higher-level patterns
"I want to understand the journey + decisions."
the-journey.md— narrativewhat-didnt-work.md— companion (the failures)top-lessons.md— synthesis
"I'm a solo developer using AI assistants and want to do this kind of work."
working-with-claude.md— the collaboration playbookthe-journey.md— applied case study (this project)top-lessons.md— synthesis
"I'm trying to navigate the codebase."
repo-map.md— folder + file maphow-it-was-built.md— what each pipeline stage does- Then dive into the GitHub repo at github.com/anandkaman/ControlMT — the actual training scripts, model code, and pipeline live there
"I'm interested specifically in the model itself, not the project around it."
You probably want the public model card instead — ../release/README.md — which has the FLORES benchmark scores, intended use, limitations, citation info. Then come back here for the journey context.
What's NOT in this folder
- Live training instructions or hyperparameters — those are in
../TRAINING_GUIDE.md(the public methodology doc) - Deployment recipes (CPU / GPU / Docker / SDK) — those are in
../release/DEPLOYMENT.md - Strategic positioning vs competitors — in my private working directory
- Old design docs from earlier phases — preserved in
archive/(10 historical docs: original architecture, training-log-v1, dataset-migration plans, etc.). These are referenced from the new docs but kept frozen.
A note on voice
These docs are written in first person, in my voice (Anand Kaman). I describe my decisions, my mistakes, what surprised me. That makes them more readable but also more vulnerable — I'm naming specific things I got wrong and what I'd do differently.
If you're an ML engineer with similar constraints (single GPU, solo, no funding), I hope the failures are at least as useful as the successes. The dead ends saved me weeks. Maybe this writeup saves you some.
Found a problem with these docs?
The repo is public at github.com/anandkaman/ControlMT. Open an issue or PR. The docs in this folder are versioned alongside the model itself — when v2.4 ships, I'll update the journey to extend through v2.4 and re-publish.