Home / Lab

Lab

Small working experiments — no videos, no mockups. Everything below runs live in your browser right now.

Demo 01 · Conversation

hire.exe

The whole pitch, command-line edition. Try roast if you're brave.

hire.exe v1.0 — the portfolio, interrogated.

Type 'help' to begin.

Demo 02 · RAG

RAG lab — bring your own key

A full retrieval pipeline in your browser: ingest text, images (OCR), and video (transcription) → smart chunking → retrieve (BM25 now, local embeddings on demand) → generate with Gemini, Anthropic, OpenAI, or any OpenAI-compatible endpoint. Your key stays in this tab's memory — it is never stored and never sent to this site.

1 · Documents (0) · 0 chunks

2 · Chunking & retrieval

Word windows with overlap.

Keyword search — instant, offline.

3 · Provider & key (memory only)

4 · Ask your documents

Load the samples, pick a provider, ask “What is Arjav skilled at?”

Demo 03 · LLM ops

CACHED RATES

Token & cost estimator

Paste a prompt, pick a model from the live models.dev catalog, budget the call before you make it. Rough estimate, not billing — but right to an order of magnitude.

MODEL · $0.15/$0.6 PER 1M

OpenAI · 128,000 ctx · tools

INPUT

~20

OUTPUT

~500

EST. COST

$0.0003

CONTEXT USED

0.4%

Heuristic: words × 1.3 + punctuation × 0.5, CJK chars × 1. Prices stream live from models.dev once loaded — showing cached rates for now.

Demo 04 · Retrieval

Similarity playground

Type a query, rank three candidates with trigram vectors. Watch exact wording win — then ask why “automobile” would lose to “car”. That gap is what real embeddings close.

#1 · CANDIDATE 1

0.403

To reset your password, open account settings and click “forgot password”.

#2 · CANDIDATE 3

0.231

Changing your password requires verifying your email address first.

#3 · CANDIDATE 2

0.023

Our refund policy covers purchases made within 30 days of checkout.

Toy model: character-trigram overlap. Try synonyms with no shared letters and watch it fail — the RAG demo above uses MiniLM embeddings precisely to survive that.

Demo 05 · LLM ops

Prompt linter

Paste a system prompt, get a pre-flight report before you spend a cent on it. Static checks — no API calls, nothing leaves your browser.

67/100

CLOSE

  • Defines a role: pass

  • Specifies output format: pass

  • Sets constraints: pass

  • Shows examples: fail

    One input→output example beats three paragraphs of description.

  • No weasel words: pass

  • Healthy length: warn

    Aim 60–2000 tokens: thinner under-specifies, longer burns budget.

Demo 06 · Text

Regex tester

Type a pattern, watch it match live — groups, flags and all. Invalid patterns tell you why instead of exploding. Nothing leaves your browser.

2 MATCHES

reach me at arjav.3003jain@gmail.com or ping admin@site.dev

Demo 07 · Data

JSON formatter

Paste a blob, get clean output plus a verdict — valid with a census, or the exact reason it fails. Your API keys stay yours: all local.

INDENT

KEYS

7

MAX DEPTH

3

OBJECTS

2

ARRAYS

2

MIN BYTES

117

{
  "model": "gpt-4o-mini",
  "max_tokens": 500,
  "stream": true,
  "stop": [
    "\n\n",
    "END"
  ],
  "meta": {
    "retry": 3,
    "tags": [
      "lab",
      "demo"
    ]
  }
}

Demo 08 · Retrieval

Chunking visualizer

The step nobody demos: how a document becomes retrievable pieces. Tune size and overlap, watch chunks form live — token counts from the same estimator as the cost demo.

2 CHUNKS · AVG ~99 TOKENS

CHUNK 1

~122 tok

Retrieval-augmented generation grounds a language model in your own documents. Instead of answering from weights alone, the system retrieves the most relevant passages and hands them to the model as context. Everything h…

CHUNK 2

~76 tok

two chunks and neither scores high enough to be retrieved. Overlap is the cheap insurance. Repeating the tail of each chunk at the head of the next keeps boundary sentences whole in at least one chunk. Typical starting p…

Rule of thumb: 300–800 tokens, 10–20% overlap — then measure hit-rate. Oversized chunks dilute retrieval; undersized ones shatter meaning at the boundaries. The RAG demo above retrieves over chunks cut like these.

Demo 09 · Ops

Cron explainer

Paste 0 2 * * *, get plain English plus the next five real runs in your timezone. Names (mon, jan) welcome.

At 2:00 — every day, every month.

+1 → Mon, Oct 5, 02:00 AM

+2 → Tue, Oct 6, 02:00 AM

+3 → Wed, Oct 7, 02:00 AM

+4 → Thu, Oct 8, 02:00 AM

+5 → Fri, Oct 9, 02:00 AM

Standard semantics: when day-of-month AND day-of-week are both restricted, a run fires when either matches. (This site's own nightly backup is 0 2 * * *.)

More experiments ship here first