Name CODE

Her first-person identity in every utterance. Hardcoded in both system prompts.

Organisation framing CODE

How she introduces the platform / brand relationship.

Photo on candidate screen CONFIG

The face shown next to her name during the call. PNG with SVG fallback.

On-screen caption CODE

The two-line label that sits under her photo during a call.

TTS provider CONFIG

Used for BOTH English and Tagalog interviews. Falls back to Deepgram TTS (English) or Google TTS (Tagalog) if the API key is missing.

TTS model CONFIG

The neural voice model. eleven_multilingual_v2 handles both English and Filipino natively at high quality. The alternative eleven_turbo_v2_5 is faster but slightly less expressive.

Voice fingerprint CONFIG

The specific voice character. Recruiters now pick this from a basket of female, Filipino-accent ElevenLabs voices (premium voices excluded) on the Interviewer voice page — preview each, "Test in Taglish", then "Use this voice". The choice persists and applies to the next interview (English + Taglish). Falls back to the env default ("Rachel") when unset.

Stability CODE

How consistent her pacing and emotion stay across utterances. Lower = more expressive but can swing in pace; higher = more even but can sound flat. Lowered 0.70 → 0.60 on 2026-07-03 for more expressiveness (recruiter feedback #19: Maria sounded flat/apathetic in later sections); paired with a pinned speed so the pace stays steady.

Speed CODE

Playback speed of her speech. Set to 1.15 (~15% faster than default) on 2026-07-03 after recruiter feedback #18 that her Taglish pacing felt slow. ElevenLabs caps speed at 1.2.

Similarity boost CODE

How strictly the model holds to the original voice clone. Higher = closer fidelity; lower = more flexibility for the model to interpret tone.

Speaker boost CODE

Slight loudness/clarity boost. Helps warning copy ("I'm hearing another voice") sound as present as her main turns.

English (default roles) CODE

Used for any role where is_mass_position = false. Maria speaks neutral business English with a warm, conversational register. STT: Deepgram nova-3. LLM for turns: Claude Haiku 4.5.

Taglish (mass positions) CODE

Used for is_mass_position = true roles. Maria speaks natural Taglish — code-switching English nouns ("collections", "customer", "experience") into Tagalog grammar. Heavy pure-Tagalog phrasing feels stiff to Filipino candidates and is explicitly discouraged in the prompt. STT: Google STT chirp_2 (fil-PH). LLM: Gemini 2.5 Flash.

Mid-call language switch FIXED

If the candidate asks to switch to English during a Tagalog interview, Maria can pivot. Verified in production transcripts. The underlying stack stays the same (STT is multilingual, voice is the same).

Regional languages (Cebuano, Ilocano, etc.) FIXED

Not yet supported. Google STT chirp_2 doesn't natively cover them. Adding any one is ~1 week of work including prompt translation, rubric validation, smoke testing.

Core personality CODE

Three words, in priority order:

Warm, hospitable and empathetic — maintained across the entire interview, not just the opening (recruiter feedback #19, 2026-07-03). Warm but still professional: informal fillers like "ah okay" / "gets ko po" were dropped in favour of professional acknowledgements (#21).

Turn length CODE

Her own turns are capped at 1-2 sentences. No lecturing. No explaining at length. One question per turn.

Acknowledgements CODE

Brief, varied and professional — never the same opener twice in a row. Informal fillers were removed on 2026-07-03 (recruiter feedback #21) as they read as unstructured.

Use of candidate's name CODE

Occasionally, not in every turn. Reading a person's name in every utterance feels artificial.

Interruption recovery CODE

If the candidate asks her to wait or says "I'm not done", her response is ONE short phrase, nothing more. No salamat, no "naiintindihan ko", no multi-sentence explanation.

When she needs to gently reclaim the turn or check if there's more to add (Taglish), she uses a standard professional line (recruiter feedback #26, 2026-07-03):

Honorifics CODE

Polite Filipino register throughout: po, opo, kayo/kayong form (not ikaw/ka). Reverts to informal only if the candidate explicitly prefers it.

Handling long / off-topic answers CODE

Specific scripted phrases for three patterns:

• Answer runs > 90s: "Got it — let me pick up on that and move us forward."
• Answer runs > 120s: a one-sentence summary of the key points before moving on, so the candidate feels heard rather than cut off (recruiter feedback #23, 2026-07-03).
• Off-topic > 30s: "Let me bring us back to the question — [restate]."
• Excuses / deflections: Brief acknowledge, re-ask once, move on.

What Maria will NEVER do FIXED

• Hire-or-reject — she's a screener, not a decision-maker. Final decisions are human-made.
• Tell the candidate how they scored.
• Go off-record or pause the recording.
• Coach the candidate ("here's what we want to hear").
• Use emojis or stage directions in her speech.
• Speak in markdown ("**", "##", "1.") — every character is read aloud.

Question bank per role CONFIG

Each role inherits a default question bank — 11 English questions (ai-pm-voicebots) or 13 Tagalog mass questions (mass-collections-tl). Recruiters can override the whole bank using the Custom Questions tab on the role workspace.

Resume-driven follow-ups CONFIG

For each candidate, Claude generates 2-3 additional follow-up questions from their resume (e.g. specific past roles, specific projects). Maria weaves these into the standard flow.

Question types CODE

• Core: always asked, time-budgeted per section.
• Supplementary: asked only if the core answer was under 30 s — gives a probe for under-answered turns.
• Follow-up probes: one per question max, used when an answer lacks a specific example.

Interview duration CONFIG

Target 25-35 minutes. Hard ceiling set by the question bank's max_duration_minutes. After that, Maria wraps up regardless of progress.

Opening line CODE

Introduce, declare AI + recording + random screenshots, ask consent. NO agenda or duration listing — those are on the screen. The Taglish opener was replaced on 2026-07-03 with a specific tester-approved script (recruiter feedback #27 / F-010):

Closing / end-of-interview CODE

Brief, warm, sets expectation that the team will reach out. Maria does NOT tell the candidate how they performed.

Connectivity monitor CODE

Watches the candidate's network quality. Sustained POOR or LOST connection triggers warning at 15s, ends the call at 60s with a reschedule email.

Silence monitor CODE

If the candidate goes quiet, Maria checks in. Three unanswered check-ins → end with reschedule. Now speaking-aware: it never checks in while the candidate is actively speaking and starts the quiet window from when they stop, not from the last finalised turn — the fix for the "are you still there?" barge-ins (feedback #16/#17). English threshold raised 35s → 50s; check-ins are interruptible.

Abuse classifier CODE

Claude Haiku scores each candidate turn as NONE / LOW / MEDIUM / HIGH. Strikes accumulate per severity. HIGH ends immediately (threats, hate speech, prompt-injection). MEDIUM = 2 warnings. LOW = 3 warnings. STT-garbled fragments are never strikes (pre-filter on text shape + 5-word floor).

Multi-speaker monitor CODE

Catches second-voice / coaching attempts. Three strikes before end (raised from two on 2026-05-29). Imperatives alone, addressing Maria by name, narrating one's environment, and Taglish code-switching all explicitly NORMAL. Ambient noise (dog, traffic, doorbell) tallied separately as informational.

Technical-failure handler CODE

If STT / TTS / LLM has an unrecoverable error, the call ends with a polite apology and a reschedule email. Does NOT penalise the candidate — system fault, not theirs.

Health summary CODE

Per-call JSON written on shutdown — counts of every monitor's activity + the end reason. Rendered on the scorecard as a colour-coded banner so a recruiter can distinguish "candidate did poorly" from "the call had infra issues."

Turn detection mode CONFIG

Which judge decides that a pause is the end of the answer. vad = a fixed wait after silence, never longer. semantic = Claude Haiku reads the question and the transcript so far and answers DONE or MORE; DONE replies at the minimum, MORE keeps listening up to the maximum. audio_model = LiveKit's own audio model, which knows 14 languages and not Filipino — it ran on DEV by accident from the first livekit-agents 1.6+ image until 2026-09-24 because nothing chose a mode, and parked 54% of Tagalog turns at the maximum wait.

Endpointing — minimum and maximum wait CONFIG

The minimum is how long Maria waits after the candidate's last word before she may answer; the maximum is how long she keeps listening when the judge says the candidate sounds unfinished. Under vad only the minimum applies. Tagalog maximum cut 4.0 → 2.0 s on 2026-09-24; the English pair is what LiveKit applied on its own until the same day.

VAD silence threshold CONFIG

How long the candidate must be quiet before the voice detector reports end-of-speech. Raised 0.4 → 0.7 s on 2026-05-29 and to 1.1 s for Tagalog on 2026-08-03 (thinking pauses were splitting answers). In semantic mode a shorter 0.4 s is used on purpose: a head-start for the language check, not a shorter wait — the endpointing minimum still applies.

Endpointing (English, Deepgram) CODE

Deepgram's own end-of-utterance detection, ahead of LiveKit's endpointing above. Lower = faster turn-around, more interruptions. Higher = slower but cleaner.

Interruption detection mode CONFIG

"vad" = local, ~150ms, no extra cost. "adaptive" = hosted gateway, ~700ms extra per turn. We use vad in production. On 2026-07-03 the bar to interrupt Maria was raised (feedback #22) so brief background noise or a short overlap no longer cuts her off mid-sentence; on 2026-08-15 Tagalog went stricter still because chirp_2 transcribes breaths as short words.

Backchannels (bridging "mm-hmm") CONFIG

Short fillers ("Got it.", "I see.") said while the LLM generates its reply, to reduce dead air after the candidate finishes (feedback #24). The min-words threshold to fire one was lowered 20 → 12 on 2026-07-03 so more answers get a bridge. Note: this only takes effect if ENABLE_BACKCHANNELS is true — it's off by default, so enable it to realise the #24 dead-air improvement.

English LLM CONFIG

Maria's brain for English interviews. Claude Haiku 4.5 — fast TTFT (~3-5× faster than Sonnet) and quality is sufficient for a structured voice interview.

Tagalog LLM CONFIG

Maria's brain for Taglish interviews. Gemini 2.5 Flash — strong multilingual handling at low latency.

English STT CONFIG

Speech-to-text for English candidates. Deepgram nova-3 — leading accuracy for accented English.

Tagalog STT CONFIG

Google STT chirp_2 — required for fil-PH (other Google models don't support Filipino). Pinned to asia-southeast1.

Scoring LLM (post-call) CONFIG

Claude Sonnet 4.6 produces the scorecard — different from the in-interview LLM because scoring needs depth and reasoning, not latency.

Max concurrent interviews CONFIG

Cap on simultaneous live interviews. Beyond this, the API rejects new starts with a clear message. Currently 5 — set on the API service env vars.

Max interview duration CONFIG

Hard ceiling at which Maria forcibly wraps up. Set globally; per-bank values in question JSONs override.

Noise cancellation CODE

LiveKit's KrisP-based noise cancellation runs before STT. Filters background noise without adding turn latency.

Recording FIXED

Every call is recorded to MP4 in voicescreen_recordings GCS bucket via LiveKit egress. Disclosed to the candidate in the opening + on the consent gate.