Skip to content
Puccha Puccha KM

Strategy — Moat & future-proofing

North star (one line)

Section titled “North star (one line)”

Each organization hires a domain professional — grounded only in that org’s own knowledge — who answers and gets things done across every touchpoint in the customer’s language, and hands off to real humans at the edge of its knowledge or authority. Never bluffing.

Multi-tenant and per-role: an org doesn’t hire one shared agent, it hires its own — isolated, scoped to its own KM — and can hire several (a rental concierge and an HR desk and a compliance desk), each bound to different touchpoints.

What this is, concretely

Section titled “What this is, concretely”
The concept In Puccha Where
“hire a professional” a Persona (role + voice + judgment + tools), provisioned from a vertical Template ADR-0167
“the org’s own KM” grounded only in Vertical knowledge (the org’s Sources) ∪ Horizontal knowledge (curated Packs), always cited ADR-0168 §2
“gets things done” Guided flows — one engine; itinerary / cart / lead / reservation / DSAR are instances ADR-0169
“every touchpoint” omnichannel; the agent emits structure, channel adapters render (widget / LINE / …) ADR-0140/0133
“hands off, never bluffs” scope-gate + citations + anti-fabrication + human handoff ADR-0147

The moat — honest

Section titled “The moat — honest”

The concept is not itself defensible: grounded agents, personas, and omnichannel are table stakes (confirmed across a 21-competitor scan — Intercom, Sierra, Glean, Salesforce, plus Thai players Amity, Botnoi, Zwiz). What competitors can’t easily hit is the specifics:

  1. Thai regulated-industry depth — 0 of 21 ship curated PDPA/BOT/SEC/OIC knowledge. Western incumbents lead with ISO/SOC2; Thai SMEs lead with nothing.
  2. Compliance-first as the whole product — never-bluff, cited, audit-logged, hashed query logs, per-tenant retention. Incumbents bolt this on; for us it is the core promise.
  3. Curated regulatory Packs — a maintained PDPA/BOT/SEC/OIC knowledge layer, grounded and citable. The one thing the whole market leaves as “bring your own.”
  4. Local SI distribution — Certogo + named channel partners into regulated Thai enterprises.

The moat is focus + localization + trust posture + distribution — not a technical wall. Defensibility depends on executing that depth (real Packs, real SI relationships) faster than incumbents care to localize.

We do NOT build a custom/fine-tuned LLM

Section titled “We do NOT build a custom/fine-tuned LLM”

A fine-tuned model bakes knowledge into weights → stale, uncitable, hallucination-prone — the opposite of the grounded, cited, never-bluff discipline that is the moat. For KM/compliance QA, grounding beats fine-tuning. “Deep-area expertise” comes from, in order of leverage:

  1. Curated domain Packs (the expertise, grounded + citable).
  2. Domain Personas (the professional’s training: voice, judgment, escalation).
  3. Retrieval + eval tuning — a domain reranker/embedding tune (small, cheap) + Promptfoo golden sets (ADR-0006) that prove quality (faithfulness ≥0.85, citation-recall ≥0.80).
  4. Model routing — Sonnet for hard reasoning, Haiku for routine; per-Persona.

Reserve model-level work for the reranker/embeddings. Consider a custom LLM only for a narrow, ultra-high-volume task where cost/latency dominates — never for knowledge depth.

Two markets, one engine

Section titled “Two markets, one engine”
  • External touchpoints — retail, tourism, sales: guided flows that convert (cart, itinerary, lead).
  • Internal compliance KM — the sharpest wedge, matching the ICP buyer (compliance officer / DPO). Staff “hire an internal professional”: upload internal policies as Sources, subscribe to regulatory Packs, provision an internal (auth-required) Persona scoped by ACL (ADR-0003), that walks staff through e.g. a PDPA data-subject request with citations and hands off to the DPO for judgment. Internal use has the lowest hallucination tolerance and the clearest compliance buyer — where never-bluff is worth the most.

Future-proof scorecard

Section titled “Future-proof scorecard”

Anti-fragile (survives foreseeable change without a rewrite):

  • Grounding-first — facts live in the org’s KM, not model weights, so model progress is a tailwind: swap/route Haiku → Sonnet → next-gen freely, cheaper and smarter, never obsoleted. The single most load-bearing decision.
  • One engine, many verticals — a new domain is data (Template + KM), not code.
  • Channel-agnostic — a new touchpoint (a new app, voice) is an adapter, not a re-architecture.
  • Compliance-first rides the regulation tailwind — rules only tighten; the product gets more valuable over time.

Execution risks (not architecture flaws):

  • The moat is focus + distribution, not a patent → defend with speed + depth of Packs + SI lock-in.
  • Agentic transactions (book/pay/provision) are deferred by design (the httpTool/action framework is deferred, not blocked) → no rewrite needed, but unbuilt; buyers will eventually expect it.

Verdict: the direction is clear and the foundations are future-proof. Remaining risk is execution — build the guided-flow engine, ship real regulatory Packs, win the SI channel — not architecture.