Journey Intelligence · sax.llc

Comprehension infrastructure
for India's DPI.

Digital public infrastructure can capture a tap in 22 languages. It cannot yet tell whether the person understood what they tapped. Journey Intelligence closes that gap — a standard for recording understanding (the comprehension artefact) and a voice engine for delivering it (Kya Likha Hai).

1 · Feel it→ 2 · Prove it→ 3 · Scale it→ 4 · Situate it
DEMO 1Experience

eKYC that explains itself

Proves: the same Aadhaar eKYC flow adapts its language, verbosity and voice to the person — and any screen can be read aloud in their language via Kya Likha Hai.

90-second script

Pick a persona and language (try Hindi). Walk the four eKYC steps. On the consent step, tap the orange KLH button and listen. Switch language mid-flow — the narration regenerates.

Open demo →
DEMO 2Record

Two records, one transaction

Proves: a consequential action can produce a DEPA-style consent artefact (permission) paired with a comprehension artefact (understanding) — generated only after a real teach-back check.

90-second script

Run “Today's flow” first — note the empty understanding slot. Switch the layer on, in Hindi. Answer the teach-back question wrong on purpose; watch it re-explain. Pass, then open “show machine record.” Fail twice: the flow pauses and writes no consent at all.

Open demo →
DEMO 3Speed & cost

Ask the screen anything

Proves: duplex voice Q&A over the same eKYC moment, answered from pre-generated cached narration (Tier 1) with live generation as fallback (Tier 2) — near-zero marginal cost for ~90% of queries.

90-second script

Pick a language, tap the mic and ask aloud — “kya share hoga?”, “wapas le sakta hoon?”. Note the response latency: that is a cache hit, not a model call.

Open demo →
DEMO 4In situ

Inside a government super-app

Proves: the whole stack generalises beyond eKYC — KLH, personas, L1–L4 materiality routing, Tier 1 cache vs Tier 2 live with latency and cost on screen, voice follow-ups, and a CAS gate that forces a review on consequential actions — inside an Umang-style government services app.

90-second script

Pick a persona (try Rajan · Farmer). Open a service and tap Compare — Tier 1 cached (instant, $0) vs Tier 2 live (Claude → Sarvam), latency and cost side by side. Ask a follow-up by mic. Then open an L4 service (NACH Mandate) and tap the red action button: the CAS gate fires with a forced 5-second review while KLH explains the consequences aloud.

Open demo →

The routing logic under every demo — transaction materiality

Full L0–L4 explainer →
L0 · NONE

Just looking. Nothing changes, nothing is spent.

L1 · INFORMATIONAL

Read-only lookup. On-device, $0, nothing to undo.

L2 · NAVIGATIONAL

Small, routine, known. Cache lookup, ~$0.

L3 · TRANSACTIONAL

Real money, larger or unknown party. Cache, then live. Hard to undo.

L4 · CONSEQUENTIAL

Irreversible, high-stakes. Max effort + CAS gate.

CAS gate — the stop sign at L4. Before an irreversible action, the system surfaces "what happens if you proceed," forces a review pause while KLH explains aloud, and requires an explicit confirm or decline. See it fire in Demo 4 (open NACH Mandate, tap the red button).

The consent artefact records permission. The comprehension artefact records understanding.
Journey Intelligence is the system that makes the second one possible at a billion-user scale.

comprehension-artefact/v0.1 draft schema · built on Sarvam Bulbul, Claude & Redis tiered caching