बकवास — Hindi/Urdu, noun. Nonsense, drivel, hot air, chitchat. The stuff you say a lot of but rarely re-read.

A browser-native explorer for your social-media archives. Started as a click-to-delete tool for Twitter/X — turned out the harder problem isn't deleting, it's finding what's worth deleting in 200,000 posts. So mostly it's now an explorer that happens to make deletion easy at the end.

Supported sources:

• Twitter / X — full surface (ZIP or unzipped folder via the File System Access API)
• Mastodon — full surface except engagement-driven cards (their export doesn't preserve fav/boost totals)
• Bluesky — live AppView API fetch by handle, no auth, engagement counts intact, full surface
• Reddit — posts + comments from the GDPR-export ZIP, no engagement counts

→ Try it live

Drop your archive (or paste a Bluesky handle), wait for the one-time index pass, and explore through seven tabs:

Counts, top hashtags / mentions / sources / languages, GitHub-contributions-style temporal heatmap spanning your full archive, hour × day-of-week heatmap, engagement distributions. Top-N panels accept a date-range filter at the top; click any heatmap day to drill into that single day's tweets in Search.

Fast full-text over tweets and likes via MiniSearch with a rich filter surface:

• Power query syntax in the search box: since:2018-01-01 until:2019-12-31 from:elonmusk to:dril min_faves:10 min_rts:5 has:media has:link lang:en -crypto "exact phrase" cricket OR football
• Visual filter bar: Since / Until / Min ♥ / Min ↻ / Language dropdown / Near match toggle / Reset
• Filter chips: Originals / Replies / Retweets / Media / Links
• Quick-range presets: 7d / 30d / 90d / YTD / This year / Last year / All
• Date-range scrubber — SVG dual-thumb range picker over a tweets-per-month histogram. Drag thumbs or the highlighted band; two-way sync with the date inputs
• Scope toggle: Tweets / Likes / Both
• Near match toggle — folds in phonetic-encoded variants so amrika/umrika find america, and Latin chirag finds Devanagari चिराग (phonetic indexing runs at MiniSearch-build time)
• Semantic search — tick Semantic to search by meaning rather than keywords. Runs all-MiniLM-L6-v2 in the browser (one-time embedding pass, results cached in IndexedDB). Large archives can use the offline Python script (scripts/embed.py) instead — 10–50× faster, uses your machine's CPU/GPU, outputs a .bin file you import once via ↑ Import emb.
• Filters-only searches — leave the text box empty, set min_faves:100 since:2018, get tweets streamed via IndexedDB cursor with the created_at index
• Select-all matches + per-result selection — adds to the global selection set

KPI cards (most-engaged month, most-engaged year, all-time top tweet, lifetime engagement), three side-by-side top-tweets lenses (combined / likes / RTs), engagement-over-time monthly line chart, viral-moments timeline, hashtags by avg engagement, sources by avg engagement, hour × DOW colored by avg engagement. Honest "views aren't in the archive" disclaimer.

Algorithmic deep analysis (no AI in the loop):

• Activity rhythm: longest streak, longest silence, monthly tweets line chart, activity change-points (≥2σ from rolling baseline), verbosity drift
• Eras compared: chronological-thirds cohort cards side-by-side (mix, avg eng, era hashtag, top hashtag, top mention, source, top tweet)
• Era detection: burst hashtags (≥50% of uses in one month) — #nft Mar 2022, #covid Mar 2020
• Mood, voice & rhythm: emoji mood timeline + word-list polarity drift (both pos − neg per month), style fingerprint with derived archetype ('Long-form · Asker · Emoji-heavy'), sleep / quiet-hours window per year
• What you talk about: top phrases (bigrams), recurring themes (bigrams in ≥3 years), hashtag-lifecycle sparklines, top link hosts you share
• Relationships: best friends (top 12 bubble cards with colour-coded initials — click to drill into a "Conversation with @user" view), distance circles (top 24 connections placed at radii inverse to interaction count, with Inner/Trusted/Acquaintances rings), top co-mention pairs, force-directed conversation graph, longest self-reply threads + a "Browse all threads →" thread explorer modal with text search and click-to-expand trees
• First, last, best: first & last original tweet of each year; top tweet per year
• On this day: today's-date anniversaries across years + "what were you doing then?" date picker with ±N day window
• Cleanup candidates: zero-engagement counts per year + sample list of 50 oldest + PII audit (regex-flagged emails / phones / SSN-shape)

Pick a year or all-time, get a poster-style summary card with top tweet, top hashtag, peak month, era hashtag, posting personality. Download as standalone HTML (hostable anywhere) or PNG (postable to social). All composed locally; nothing uploaded.

Hashtag co-occurrence clusters with an adjustable threshold. Pick a theme, drill into it via "Search this cluster".

The deletion confirmation tray:

• Per-card actions: open in X, mark deleted, undo
• Bulk actions: "Open N tabs" (modal guard at >10, hard cap at 50), "Mark N deleted"
• BYOK Twitter / X API delete — paste your own 4 OAuth 1.0a credentials, Bakwaas signs DELETE /2/tweets/:id with HMAC-SHA1, handles 429 rate limits with reset-aware backoff, marks deleted in the local store automatically
• Site-wide select buttons — every tweet card in every tab has a small + select button. Counts roll up into a floating chip visible from any tab

Click-to-delete is honest (the default). Bakwaas opens the tweet on X in a new tab; you delete it there in the real UI; you come back and confirm. The hard part — knowing what to delete — happens in Bakwaas. The actual delete happens in X's own UI.

BYOK X API delete (optional). If you have your own X developer credentials (OAuth 1.0a Read+Write), Bakwaas can bulk-delete via DELETE /2/tweets/:id directly. Keys live in IndexedDB only — never the FSA sidecar, never anywhere except api.x.com. Honest about cost: X v2 DELETE is a paid endpoint these days.

No X API. No OAuth. No accounts. No server. Every export is a download from your browser, not an upload from it.

One transparent exception: Bakwaas runs Google Analytics 4 (G-E5XCNWFXNC) for page visits and tab opens — a basic "is anyone using this" counter for the project. No archive content is ever sent, only the URL and which tab you opened. anonymize_ip is on. Block it with uBlock / Privacy Badger if you prefer.

Everything runs locally. The archive never leaves your tab. The Bluesky path makes only the API calls you'd make yourself. The X API delete path only ever calls api.x.com. No NakliTechie server; no telemetry; no logging.

Where state lives:

• IndexedDB (origin: bakwaas.naklitechie.com) — the parsed index, MiniSearch serialized, all aggregation caches, BYOK X credentials, deletion state
• OPFS mirror at bakwaas-state.json — backup of phase / fingerprint state
• FSA sidecar at <archive>/bakwaas-state.json — folder-mode only, never carries BYOK keys (sidecar travels with the archive)

• Single-file HTML, vanilla JS, no build step. Open with python3 -m http.server or any static host.
• MiniSearch for full-text (CDN ESM @ 7.1.2). Indexes serialised to IndexedDB. Lazy rebuild on first search.
• JSZip for ZIP ingestion (CDN ESM). FSA folder mode skips it entirely.
• Palette: Rangrez — periwinkle dark (--accent:#6b8ef5, --accent-2:#f0845a coral). Light mode via .theme-light. Toggle in the header persists to localStorage; auto-detects OS preference on first visit.
• Phonetic indexing: per-tweet phonetic_text field generated at index time. Cross-script transliteration covers nine non-Latin scripts — Brahmic family (Devanagari, Bengali, Tamil, Gurmukhi, Gujarati, Telugu, Kannada, Malayalam) via shared inherent-a/halant logic plus Cyrillic and Greek as alphabetic 1:1. CJK covered: Chinese → pinyin, Japanese kanji → kana → romaji, Korean → romaja. Latin output runs through a consonant-skeleton encoder. Result: chirag, chiraag, and चिराग all hash to krg; america/amrika/umrika → mrk.
• Semantic search: @xenova/transformers@2.17.2 / all-MiniLM-L6-v2 runs on the main thread (384-dim vectors, chunked with setTimeout(0) yields). Embeddings stored in IndexedDB (DB_VERSION=3, embeddings store). For large archives, scripts/embed.py generates the same vectors offline via sentence-transformers (batch_size=512); output is a BWEM binary (magic + uint32 header + null-separated ID block + raw float32 matrix) imported once. CSP must include https://hf.co https://*.hf.co https://*.xethub.hf.co — HuggingFace redirects large ONNX files to the XetHub CDN, not huggingface.co.
• OAuth 1.0a signing for the X API delete path: HMAC-SHA1 via SubtleCrypto, RFC-3986 percent-encoded base-string assembly.
• CSP pinned: default-src 'self', script-src 'self' 'unsafe-inline' https://cdn.jsdelivr.net, connect-src extended for BYOK provider hosts + HuggingFace / XetHub CDN for model weights.

What's in PENDING.md (local-only) — the next-session work:

• Semantic "Find more like this" — surface the top-K semantically similar tweets on any result, now that embeddings are cached in IDB
• Topics cluster labels — derive 2–4 word labels from centroid embeddings per hashtag cluster (no LLM needed)
• Arabic phonetic indexing — different structural pattern from Brahmic (right-to-left, short vowels typically omitted); Buckwalter-style consonantal skeleton
• Instagram archive as a fifth provider
• BYOK LLM (Anthropic / OpenAI / Gemini) — natural-language → filter spec in Search; cluster labels in Topics; bulk classify in Patterns

Bakwaas is a sibling of a handful of other browser-native, single-file tools by the same author. All zero-server, all your data stays on your device.

See the full series →

Built by Chirag Patnaik Built with Claude.