Intervues
◆ Public diligence packet · no pricingWhat was actually hard to build

The parts that were
hard to build.

Not a feature list — the actual mechanisms underneath the voice interviewer: a fused turn-taking model that proves it's safe before it interrupts you, a memory engine written in Rust from scratch, and a report generator that deletes its own claims rather than invent a citation. Every number below is either a gate enforced in CI or labelled as a target.

Turn-taking

A 50ms fused feature stream, not a silence timer.

Most voice agents guess when you're done talking from silence alone. Ours fuses 13 signals every 50ms — VAD, prosody, partial-ASR stability, and semantic completeness — into an endpoint predictor that outputs three separate calibrated probabilities, each Platt-scaled against labelled data:

p_turn_end
Commit threshold ≥ 0.80 before the room replies
p_backchannel_safe
Gates whether a “mhm” is safe to speak
p_co_completion_safe
Gates whether finishing your sentence is safe
The false-interrupt gate

Backchannels don't ship until they prove they're safe.

At boot, the system replays a labelled mock-interview corpus through the endpoint predictor and measures how often it would have falsely interrupted a candidate who was still thinking. That rate has to clear a gate before backchannels are allowed to speak at all:

the gate

< 2% false-interrupt rate

Measured on labelled HOLD points from the mock corpus. Above the gate, or if the corpus is missing, or if the eval itself errors — the mode pins to shadow, not live.

Fail-safe by design
what shadow mode means

It computes the decision. It doesn't speak it.

Every backchannel and co-completion decision is logged and scored against ground truth in production, silently, until the measured rate earns the right to go live.

Default posture: prove it first
Indian-English calibration

A named, versioned control — not a vibe.

The five-tier silence ladder runs on pause-tolerance tiers multiplied by a single documented constant, so the calibration is something an auditor can actually point at:

×1.25
INDIAN_ENGLISH_MULTIPLIER — applied to every pause-tier boundary before the ladder decides to reassure, offer agency, reclaim the turn, or redirect.
5
Ladder tiers— HOLD → REASSURE → OFFER_AGENCY → RECLAIM → REDIRECT, each with 0.15s hysteresis so the system doesn't flicker between tiers on a single noisy frame.
Caged sentence completion

It may finish your sentence. Never your numbers.

Co-completion is only allowed to close safe grammatical tails — and is hard-blocked by regex from ever completing anything that matches a number, an architecture term, or a scored claim. Two wrong completions in a single session disable it entirely, for that candidate, for the rest of the interview.

Memory

A hand-rolled vector engine, written in Rust.

Not a wrapped vector database — a purpose-built quantization scheme (“TurboQuant”): a seeded orthogonal rotation, a Lloyd-Max 3-bit codebook, a 1536-bit QJL residual sketch, and a two-stage scan that reranks the sketch's shortlist against full fp16 vectors. Served over gRPC, contract frozen and additive-only since 2026-05-21.

4
Tiers — working set, brute-force/TurboQuant, disk shards, pinned atoms
≥0.95
recall@8 CI gate — the build fails below it
≤70ms
Tier-2 retrieval p95 CI gate
Latency as a source of truth

Eight named stages. One budget. Enforced, not aspirational.

Every turn is traced through audio capture, ASR finalization, endpoint commit, answer embedding, memory recall, context assembly, the next-question model, and first-audio TTS — each stage flagged the instant it goes over budget, with live p95 queryable in production.

5420ms
Sum of all eight stage budgets, per turn
5500ms
Hard gate — the ceiling the budget isn't allowed to cross
900ms
Target for first audio out of text-to-speech
Reports that can't hallucinate

A citation your report can't verify never reaches you.

The report drafts in one JSON-mode call, deliberately separate from the latency-tuned interview brain. Then every cited turn is re-resolved against the actual transcript: a citation to a turn that doesn't exist is dropped, and a claim left with zero valid citations is deleted from the report entirely — not flagged, removed.

Built to be handed over

Not a demo. A codebase with its own eval harness.

~130 / ~130
Source files to test files in the realtime service — roughly one test file per source file, not a bolt-on suite.
5
Fault-injection toggles — ASR, TTS, the brain, the embedder, and memory can each be forced to fail so the fallback paths get exercised on purpose, not by accident.
An A/B harnesscomparing an India-first model against a frontier model on the same labelled corpus, plus criterion benches on the memory engine's hot paths.
🟡/⚪
A feature matrix, not a launch page — every capability is marked proven-offline, needs-live-verification, partial, or flag-off. Co-completion and model routing are flagged off by default until their own gates clear.

This is the part most pitches skip.

No pricing here, no pilot to book — just the mechanisms. If you're evaluating the engineering directly, the build log is in the journal, and a real inbox is on the other end of the email below.

Start practising →