Bhaify

Not a chatbot. One box, one button, one line back — your message rewritten in a very specific voice, as an openly-labelled fan parody. It never answers you, never converses, never adds a claim of its own.

The persona is the product, so every interesting problem in it turned out to be a measurement problem.

The corpus comes first

52 verbatim tweets, 2010–2026, extracted from a fan site’s embedded RSC payload by script rather than retyped.

• Three records came back clipped — two contained a " that killed the extractor’s string match, one was truncated by X’s own “Show more” and had to be recovered from press coverage.
• verify-corpus.ts accepts a recovered record only if the scraped text is a strict prefix of it. A recovery can extend a record, never quietly rewrite one.
• Grounding data only. It feeds few-shot selection and eval scoring; no page ever renders it.

The design is the joke

Borrowed x.ai’s language wholesale — near-black canvas, white outline pills, one geometric sans at weight 400, uppercase tracked mono labels. Its own design doc calls the effect “a research lab announcing its work rather than a SaaS marketing site.”

The voice it wraps is the opposite: loud, filmy, maximalist. Play the container completely straight and the content does all the work. A Bollywood-gold treatment would be more obvious and much less funny.

The eval harness was certifying the failure

It reported 28 of 31 passing on output that felt nothing like him.

• A “style marker” was a match on /aa|oo|uu/ or the token " ." — so pure Hindi transliteration scored for free, and a three-beat line banked three markers before saying anything.
• “Has a structural move” passed on segments >= 3, which is the three-beat template it was supposed to catch.
• A 0.5×–2.5× length band enforced the clipped rhythm that made everything read like a Mad Lib, against a corpus that rambles to a median of 18 words.

It was measuring fidelity, which was already fine, and calling it persona.

The rebuild splits them into two axes that fail independently:

Axis	Scope	Checks
Fidelity	per output	a question stays a question; names and numbers survive; no assistant-voice tells
Style	per batch	compression, English-base share, rhythm, median length, sign-off rate, phrase repetition

Style has to be per batch because inconsistency is the style. One line that compresses nothing is fine; a batch where nothing compresses is broken.

One metric earns its place by refusing to average — pureHindiShare. A batch averaging 9% English isn’t 31 lines at 9%, it’s a third at 30% and a third at zero. The zeroes are the broken output hiding behind the good ones.

What the numbers found

• The few-shots were teaching the wrong task. 1.4 SMS tokens per 100 words and 0.2% English function words, against a corpus at 8.5–13 and ~19%. Fifty hand-written examples were demonstrating translate this to Hindi.
• The style guide was decoration. It asked for sign-offs under 10%; the examples used them 20% of the time and live output hit 45%. New rule: anything asserted in style-guide.ts must be visible in examples.ts.
• The deterministic quirk pass was inert. Its lexicon keys on English function words and the examples were teaching Hindi, so nothing matched — quirk density was 0 on 18 of 29 live rows, and raw-vs-quirkified metrics were identical on every axis. Surface polish can’t fix a voice; it can only finish one.
• Two transforms read as bugs, not voice. The corpus capitalises mid-sentence under rhetorical stress. A uniform per-word probability can’t tell stress from filler, so production produced “Roz rath ko Loud music”.

Choosing the slower model on purpose

12 cases, variant shapes rotated:

Model	Compression	English-base	Pure Hindi	Gate	Latency
gemini-flash-lite	3.4 /100w	5%	50%	fail	1.1s
gemini-flash (capped)	13.0 /100w	19%	0%	pass	9.3s
corpus reference	8.5 /100w	19%	—	—	—

flash-lite doesn’t just score lower — half its output is a Hindi translation, the exact failure that made the product feel wrong. So the primary is ~8× slower and 503s on 5 of 12 calls. 503 is retryable, so the chain degrades to the fast one automatically.

Capping the reasoning budget is what made it viable, and the working value isn’t the obvious one: 'none' → HTTP 400, 'low' → 98 tokens to emit “ok”, 'minimal' → 8, omitted → 117.

Smaller decisions

• Variant slots. A response cache and randomised style cancel out if the key is just the input hash. Three slots per key instead; regenerate fills the next empty one, then cycles them free. Each slot binds a different shape — run-on, one-liner, pivot — because seed variation alone produced three rewordings of the same sentence.
• No AI SDK. One non-streaming call that must return an id synchronously. Gemini’s OpenAI-compatible endpoint means both providers take an identical payload, so the fallback chain is a loop over a table.
• Injection defence isn’t a coin flip. User text sits inside <message_to_bhaify>, with the closing instruction after it. Two adversarial few-shots are pinned into every request — if the sampler could drop them, it would drop them on the request where they matter.
• Share cards stay plainly ours. A pixel-perfect fake tweet screenshot turns parody into a fabricated quote. Own chrome, a BHAIFIED stamp, disclaimer rendered into the image — OG cards travel without their page.