I build offline speech and language AI (ASR, TTS, voice cloning, translation) for a low-resource language, solo, on consumer hardware. These are the field notes: what broke, what measured well, and what transfers.

• The guarantee lives in the execution layer
  2026-08-20
  An NL-to-SQL system over records that are evidence: why the write-guard is a read-only connection instead of a SQL parser, what that does to the attack checklist, and being honest about the approval step.
• Your agent's tools survive 4-bit. Its token budget and its defaults don't.
  2026-08-11
  300 frozen tool-calling items against five quantized variants: structure never breaks, digit copying is perfect, and the damage hides in a token budget and a class of invented arguments every validator waves through.
• NVFP4 on a 5090 names two different machines
  2026-08-07
  Same checkpoint, same card, two kernels, different numbers. What quantization actually costs on consumer Blackwell, and why no benchmark should ship without its kernel receipt.
• The analyzer disagreed with the generator, and it was right
  2026-08-06
  My lesson generator wrote a sentence. My morphological analyzer read it back as something else. Auditing the disagreement shrank my verb list from 58 to 11, and I trust the 11.
• An LLM ensemble took 59 minutes to miss a bug that regex found in 0.03 seconds
  2026-08-03
  For one very common class of continuity bug, two local LLMs and an hour of compute lost to 840 lines of standard-library Python. The lesson generalizes.
• Ask an AI to describe a writer's voice and it will describe itself
  2026-07-27
  I asked a model to describe a novelist's voice. It gave six confident signature constructions. A grep of 414,000 words of her actual prose found every one of them exactly zero times.
• Same checkpoint, two answers: a silent PyTorch bug on Blackwell
  2026-07-23
  For three weeks the recipe 'didn't work.' Every hypothesis about the training was wrong, because the training was fine. The toolchain was silently miscomputing on new silicon and the loss curves never let on. Here's the bug, the one diagnostic that caught it, and the environment-acceptance test I now run before trusting any new box.
• Round-trip CER: shipping TTS in a language you cannot listen-check
  2026-07-14
  The quality gate I use for text-to-speech when no human can proofread the audio: synthesize, transcribe, score, gate.
• The judge was stale: how a frozen evaluator almost inverted a licensing decision
  2026-07-14
  Every quality score in my TTS pipeline comes from a model. Models age. Mine aged silently and nearly flipped a real decision.