Skip to main content
SGBuilt in Singapore for Singapore SMBs · EDG-aligned

RELATIONAL AI THAT
COMPOUNDS. EDG-CLAIMABLE.

Custom relational AI agents — companions, characters, concierges, brand voices — built around your product and your customers. Powered by Gemini 3.1 Flash Lite at ~1/20th the output-token cost of GPT-5.5 or Claude Opus 4.7, sharper every quarter, and eligible for up to 50% subsidy under the Enterprise Development Grant.

~5 min · downloadable PDF brief · email-gated

S$70M+
AI value delivered at DBS
First
gen-AI rollout at OCBC
≤200ms
context retrieval p95
Up to 50%
EDG-claimable
The cost story

Frontier-quality output. ~5% of the API cost.

Most agent work is context-bound, not reasoning-bound. Sonzai is built around Gemini 3.1 Flash Lite — the Relationship Layer assembles memory, personality, mood, relationships, and a knowledge graph into every prompt, so a $1.50/M-output model lands as well as a $25–30/M frontier model on the work that matters.

Output token cost — what you actually pay
USD per million tokens · public API rates · May 2026
Live rates
GPT-5.5
$30/M
Claude Opus 4.7
$25/M
Claude Sonnet 4.6
$15/M
Gemini 3.1 Pro
$12/M
Gemini 3.1 Flash Lite — Sonzai default
$1.50/M
Your unit economics shift when retrieval, personality, and memory do the heavy lifting — not the LLM token bill. The Sonzai harness ships with Gemini 3.1 Flash Lite by default; BYOM for everything else.
~95% saved
Built around Gemini 3.1 Flash Lite

$0.25/M input, $1.50/M output. Sonzai's harness pairs it with personality, memory, mood, relationships, and a knowledge graph — so the output quality matches frontier models on the work that matters.

Switch models any time

Better lightweight model drops next month? Point a config flag at it. Your agents inherit it instantly — no retraining, no migration, no relationship reset.

BYOM for compliance

Bring your own model — Llama, Qwen, DeepSeek, fine-tunes — self-hosted via vLLM, Ollama, or TGI for regulated workloads. Same Relationship Layer, your hardware.

Compounding ROI

A HUMAN HIRE PLATEAUS.
AN AI EMPLOYEE COMPOUNDS.

Every session ends with sessions.End() — kicking off fact extraction, personality drift, mood update, quality scoring, and per-pair retrieval re-tuning. The agent that meets your customer next quarter is materially better than the one that met them today.

Effective output per S$1,000 spent
Indicative · 12-month horizon · per-pair learning loops on
Role replaced≈ 3–4 months

Lead Qualification SDR

Headcount cost · S$50–70K / yr

  • Inbound enrichment + scoring
  • Discovery calls qualified to BANT
  • CRM hygiene + handoff notes
  • Persistent context per lead
See deployment
Role replaced≈ 2–4 months

Customer Support Tier-1

Headcount cost · S$36–50K / yr

  • 70% of repeat tickets fully resolved
  • Context across every prior conversation
  • KB writes verified by the agent itself
  • Escalation packets prepped for Tier-2
See deployment
Role replaced≈ 3–5 months

Travel Ops / Itinerary Build

Headcount cost · S$40–60K / yr

  • Itineraries assembled from supplier data
  • Margin checks + auto-rebooking flags
  • Group coordination across travellers
  • Compounds with every season
See deployment

Salary ranges from public Singapore market data (MyCareersFuture, Glassdoor SG, market reports). AI agent cost depends on integration scope; typical first build runs S$30K–80K of which up to 50% may be EDG-claimable.

Why ours compound · Multiplayer Memory

One world.
Every character. Every player.

The reason a Sonzai agent gets sharper with tenure: every agent on your project shares a typed knowledge graph. One agent learns the user's preference, another picks it up on the next channel; a brand voice updates the customer context, the concierge sees it next turn. Documents auto-ingest, agents auto-update the KB, newer information supersedes older facts. Compounding context — the relational memory that doesn't reset.

  • Intra + inter-agent — same graph, all your agents share it
  • Typed facts (Person, Account, Decision, Promise, Document)
  • Documents auto-ingest; agents auto-update the KB
  • Newer information supersedes older facts, automatically
agents · users · documents · facts
session.End() · per-pair learning · evolution
Self-Learning Layer · the compounding curve

Sharper every quarter.
Without re-training a thing.

Every conversation ends with sessions.End() — kicking off fact extraction, dedup, personality drift, mood update, diary, and quality scoring. Per-(agent, customer) SGD and bandits tune retrieval to that specific account. The agent that walks into next quarter's conversation is measurably sharper than today's, with zero training infrastructure on your side.

  • Per-(agent, user) retrieval tuned by SGD on each session's reward
  • Multi-armed bandit (Thompson sampling) over memory clusters
  • Shadow-rollout RL with auto-revert — production never sees a bad policy
  • Daily/weekly jobs: tree pruning, decay, diary, milestone detection
Per-(agent, user) sharpnesscompounding
Day 1
Week 1
Month 1
Year 1
Enterprise Development Grant

Up to 50% covered.
We help you draft the case.

Custom AI agent builds typically fall under EDG's Innovation & Productivity pillar — Automation, Process Redesign, or Product Development. We have delivered AI work at DBS and OCBC, and we structure every engagement to give Enterprise Singapore the artefacts it expects: scope, deliverables, KPIs, and a final report.

01

Discovery

Free 30-min call. We confirm fit and EDG-shape.

02

Proposal

EDG-aligned scope, KPIs, milestones, and pricing.

03

Build & claim

We deliver. You submit the final report and claim.

Read the full EDG guide
EDG eligibility — quick check
  • Registered & operating in Singapore
  • ≥30% local equity (SG / SG PR)
  • In a financially viable position
  • Ready to start within ~3 months
  • Project drives capability, productivity, or revenue
Final eligibility is determined by Enterprise Singapore. We help you build the case during the discovery call.
What happens after the call

Discovery to production.
Typically 4–8 weeks.

Week 0

Discovery

30-min call, free. Workflow audit, AI agent shape, EDG-fit confirmation.

Week 0–1

Proposal

EDG-aligned scope, KPIs, milestones, integration plan, and pricing.

Weeks 1–3

Build

Relationship Layer agent design, knowledge ingestion, integrations, eval rubric.

Week 4

Pilot

Live in your stack against a controlled cohort. Quality scorecards. Adjust.

Month 2+

Scale

Production rollout, retrieval re-tunes per pair, knowledge graph compounds.

Singapore Q&A

THE QUESTIONS
FROM SG FOUNDERS.

Is Sonzai EDG-eligible?

EDG eligibility is determined by Enterprise Singapore on a per-project basis, not by us. Custom AI agent builds typically fall under the Innovation & Productivity pillar (Automation, Process Redesign, Product Development). Standard SME support is up to 50% of qualifying project costs; higher quanta are case-by-case under the Sustainability overlay. On the discovery call we confirm shape and help you draft the case.

How is this different from PSG-listed AI tools?

PSG (Productivity Solutions Grant) covers pre-approved off-the-shelf software with fixed scope. EDG covers custom development and consulting — which is what a tailored AI agent actually is. PSG is faster but rigid; EDG is bespoke. Most ROI-positive AI agent deployments are EDG-shaped because they integrate into your specific workflow.

How can lightweight models match frontier quality?

Most agent work is context-bound, not reasoning-bound. Sonzai is built around Gemini 3.1 Flash Lite — at $0.25/$1.50 per million input/output tokens, ~20× cheaper on output than GPT-5.5 ($30/M) or Claude Opus 4.7 ($25/M), while inheriting the Relationship Layer's persistent memory, personality, mood, knowledge graph, and relationship state in every prompt. For regulated workloads, BYOM (Llama, Qwen, DeepSeek, fine-tunes) self-hosted via vLLM, Ollama, or TGI — same Relationship Layer, your hardware.

Where is data hosted? Is it PDPA-compliant?

We're a Singapore-incorporated company (Sonzai Labs Pte Ltd) and operate under the PDPA. Data residency, retention, and processing are scoped per engagement and documented in your project SOW. For sensitive deployments we support private regions and on-premise integration patterns.

What's the typical engagement timeline?

A typical custom AI agent runs 4–8 weeks from discovery to production pilot. Discovery and scoping (week 0), build (weeks 1–3), integration and pilot (week 4), scale and optimise (month 2+). EDG approval typically adds 6–12 weeks before kickoff.

Do you support multilingual deployments — Mandarin, Bahasa, Tagalog?

Yes. The Relationship Layer is model-agnostic — pick the LLM that handles your target language best. We've shipped agents that operate across English + Mandarin + Bahasa across SEA. Personality, memory, and relationship state all carry across languages for the same user.

What's the smallest engagement you take?

Discovery calls are free. Pilot builds typically start at the EDG project floor (~S$30K) so a 50% claim makes the maths work. We will tell you on the discovery call if managed-build is overkill and you'd be better served self-serving on platform.sonz.ai.

YOU'RE ONE CALL FROM
A COMPOUNDING TEAMMATE.

30 minutes with the founders. We confirm if it's a fit, what it's worth to your business, and whether EDG covers half. You walk away with a plan either way.

Or book a discovery call

5 minutes · downloadable PDF · we get a copy too

Sonzai Labs Pte Ltd · Registered & operating in Singapore