Internal research projects

Key features

• Range from four recordings — read a sentence, hold a comfortable note, sing lightly, and glide like a siren from low to high. Range is measured only from sung prompts, because speech makes the range look lower and narrower and the song shrinks. The measured range (e.g. D3–A4) is shown right after recording; recordings that are too short or unsteady are retaken.
• One song, several singers — lyric sections are assigned to participants, and each person's converted vocal is joined with short crossfades at section boundaries; sections sung together are layered.
• Local-LLM lyrics and song design — writes lyrics from title, mood and genre, or polishes the customer's own. A small model (gemma-4) goes first, falling back to 27B. Nine song languages; tempo and length default to “auto” with only a few easy choices. Hip-hop and rap get their own length calculation from syllables per bar.
• Keeping notes in range — rules such as the participant's range, a one-octave span and no leaps beyond a fifth go to both the music model and the design LLM. The generated vocal's range is measured and shifted by octaves; if it still exceeds the range by more than five semitones, or holds long high notes or jumps 12 semitones, it is regenerated. Jumps are checked again after conversion.
• Lyric check — each take is transcribed with Whisper and compared with the ordered lyrics against a per-language threshold; spellings that sound alike are treated as equal. Word timings also drive the Shorts captions and the score.
• Deliverables — full track, instrumental, lyrics, and a lead sheet PDF with lyrics transcribed automatically from the AI guide vocal (LilyPond), via email links and “My music”.
• 30-second Shorts — picks a highlight by vocal presence and energy and renders a 9:16 video with line-by-line lyrics. Requires consent from every participant at order time and is uploaded to the official channel only after the orderer confirms the song; it can be withdrawn at any time.
• Revisions and retention — up to three revisions within 14 days of first delivery. Original recordings are stored encrypted and deleted on confirmation, when revisions run out or when the period ends; intermediate files are cleaned right after delivery.
• Fixable failures — lyric problems let the customer edit (or AI-polish) the lyrics and remake, recording problems allow a retake, and system problems are retried automatically twice after a few minutes.
• Payments and tax — Toss Payments in KRW and USD (international cards · PayPal). If the approval response is lost, the order stays “confirming” and the worker settles it by querying the provider. Webhook bodies are never trusted — state is re-fetched — with partial/full refunds and a ledger classified as taxable or zero-rated.
• Secure defaults — CSP and frame blocking, upload magic-byte checks and blocking of disguised playlists, order keys only in headers, admin console returns 404 outside the VPN range. Visitors from abroad land on the English UI.
• Sharing one GPU — when an order arrives the assistant's 27B is unloaded and the music engine leased; a lease processes all queued jobs, and prep runs at most two jobs ahead. Measured: about 7 minutes for a 64-second song with two singers, lyric match 0.975.
• Scaling with demand — additional LLM servers are planned as customer order volume grows. Job pickup and payment updates are serialized with database locks, so adding machines never produces the same order twice.

Architecture · one order

Key capabilities

• Three response paths — the small router answers directly or runs a single lookup tool only when it is confident (0.8+), in a few seconds to about 20 seconds; everything else goes to the 27B ReAct loop (up to 5 steps). An answer written without tools for a request that needed them is discarded and retried.
• Button-free real-time voice — it listens continuously, answers when you finish, and stops when you cut in while it is speaking. Backchannels like "uh-huh" are told apart from interruptions so it keeps talking, and an interrupted answer is remembered only up to the part actually heard. Measured: first audio about 2 seconds after you stop, and it stops within about 0.5 seconds when interrupted.
• Listens only to you — enroll about 45 seconds of voice and it accepts only registered speakers. Profiles are encrypted locally, and it learns your voice and pronunciation habits (e.g. terms it mishears) over time. Speaker verification is never used as authorization, so voice mode is lookup-only — no orders or sending.
• Says only what it verified — for live questions such as weather, stock prices or calendar it covers the wait with "let me check", then answers only from the lookup result; if the lookup fails it does not make up numbers. Stock questions all go to one briefing tool where the server collects, computes and drafts, reducing tool-selection mistakes.
• An answer within the time limit — when time runs short it stops calling tools and writes from what it has, and if that is empty it writes a partial answer from successful lookups. Runs are detached from the connection, so they finish even if the phone disconnects, and reconnecting resumes them.
• Long-term memory — durable facts (names, preferences, schedules) are extracted into a vector DB, and only those above an empirically tuned similarity threshold are recalled. Near-duplicate memories are not stored twice.
• Speaks first — a morning briefing of mail, calendar, tax deadlines, reminders and holdings (date and weekday phrases verified and corrected in code), an hourly digest of new mail, and D-day reminders synced with Google Tasks.
• Business assistant — record transactions by voice, receipt photo or card-statement file (VAT back-calculated), issue quotes and invoices with receivable alerts, store and semantically search documents, net worth and monthly reports.
• Market data and strategy validation — collects daily prices for every listed stock for screening and backtests. Backtests use only information up to the previous day and assume the unfavorable case when intraday order is unknown. A new setting must beat the current one in all four test windows to become a candidate, accumulates out-of-sample results through daily shadow runs with no real orders, and is adopted only by a human. Orders placed through conversation always require direct human approval, regardless of auto-approve settings.
• Lyrics, composition and audio — describe the song you want and the 27B model drafts a song blueprint (JSON) with sections, chord progressions, BPM and key, writes the lyrics, and the music model ACE-Step 1.5 renders a track with vocals. Blueprints are schema-validated, and truncated or malformed responses are re-requested with the reason attached. Measured: the audio-generation step takes about 45 seconds for a 30-second track.
• Sharing one GPU by swapping — 32 GB of memory cannot hold the 27B model and the music model together, so the 27B is unloaded only while music jobs run and restored afterwards; chat keeps working on the small-model path in the meantime. Generation runs in a night batch (22:00–05:30) by default; daytime requests get their design and lyrics done, then step out of the queue so they don't block other jobs. [Generate now] runs one immediately.
• Composing from reference tracks — give several YouTube links with weights and it analyzes each one's BPM, key, chord progression, structure, energy and timbre into a target profile, then designs a new chord progression and structure to match it, without reusing the references' melodies or progressions. The finished track is re-analyzed with the same analyzer and flagged if it is too close to a reference (a screening aid for possible copying, not a legal judgment). References are kept for personal analysis only.
• Sung in your own voice — after recording consent that the voice is yours or used with permission, enroll about 30 seconds of voice; it separates the vocals from a generated track or a song you own (Demucs), converts the timbre (Seed-VC) and remixes them with the instrumental, with key shifts up to ±12 semitones. Enrolled voices are encrypted and stay local, and results are for personal listening.
• Gets faster on its own — every night it diagnoses the previous day's latency, tool failure rate and negative feedback, and learns router examples from simple questions that were handled slowly so they take the fast path next time. Changes roll back automatically if eval-set accuracy drops; it never edits its own code.

Architecture · chat

Processing flow · real-time voice

• Mic stream16 kHz · 20 ms frames
• Voice detectionsilero VAD
• Speaker checkenrolled voices only
• RecognitionWhisper · pronunciation fix
• Turn detectionendings · silence rules
• Responsegemma-4 · lookup tools
• Sentence-level TTSplays from first sentence

Processing flow · music production

• Requestdescription · references · voice
• Reference analysislibrosa · tempo·key·chords
• Song design27B · sections·chords·BPM
• Lyricsper-section lyrics
• GPU swapunload 27B · night batch
• Audio generationACE-Step 1.5
• Voice conversionDemucs · Seed-VC
• Output checksfile checks · originality

Code · statisticsLLM judgmentValidation gate

※ The finance features are a technical reference, not investment advice or a solicitation to trade, and no returns are guaranteed.

Processing flow

• Collectnews RSS · DART · KIS
• Normalizecanonical schema · matching
• Extractrule parsers + local LLM pool
• Ontology check20 events · 10 relations
• Promotededupe · evidence
• RAGbge-m3
• Market reactionabnormal return · flows
• Intelligencecited insights
• Reportdaily · weekly

code · statisticsLLM judgmentvalidation gate

Architecture · promotion path

※ Technical reference. Not investment advice or a solicitation to trade; no returns are guaranteed.

• Bars · tickers · WS1-minute bars → DB · Redis
• Market-state classificationregime labels
• Judgmentrules + LLM
• Safety switchsingle LIVE flag · position cap
• Executionpaper / shadow / live

※ Technical reference. Not investment advice or a solicitation to trade; no returns are guaranteed.

Architecture

• Telegram botcoaching · analysis · payments
• FusionAgentmodel delegation
• Signed tokenephemeral · stateless
• Read-only queryno DB write privilege
• Server-side SVGrelationship network
• WebApp / PCper-page CSP