Stage 1
Launch
A deterministic platform with MVP models behind confidence gates.
- Extraction and classification
- HS top-three suggestions
- Screening and entity resolution
- ETA and demurrage risk
- Exception ranking
- Load planning and geo-tracking
Architecture · GTM
AI-native. Deterministic where it counts.
Every process is owned by a durable workflow. Every model sits behind a typed contract with a confidence value and a fallback. Every decision is traceable to a rule, a version and an author.
PRIMARY DEPLOYMENT BOUNDARY
India-primary cells
AWS ap-south-1 · 3 AZ
Execution architecture
One controlled path from request to evidence.
Select a layer to inspect its responsibility and controls. The arrows describe dependency direction; state changes still enter through explicit commands.
Cell boundary: India-primary cells · AWS ap-south-1 · 3 AZ
Architecture principles
Nine choices that keep the platform accountable.
Each principle reduces hidden state, uncontrolled coupling or ambiguity at an operational boundary.
- 01
Deterministic orchestration
Durable workflows—not prompts—own state, retries, timers, approvals and recovery.
- 02
Modular monolith, enforced seams
One deployable core keeps transactions coherent while build rules stop modules reaching across boundaries.
- 03
Two first-party BFFs, one partner surface
Web and mobile get purpose-built interfaces; partners use a separately governed API gateway.
- 04
Tenant context is infrastructure
Identity, data access, events, caches and audit records all carry tenant scope by default.
- 05
ML behind typed contracts
Models return bounded fields, confidence and provenance; callers choose acceptance, review or fallback.
- 06
Events out, commands in
Modules accept explicit commands and publish facts, preventing consumers from mutating another module’s state.
- 07
India-primary cells
The intended primary footprint is AWS ap-south-1 across three availability zones, with residency designed into cell boundaries.
- 08
Privacy as infrastructure
DPDP rights, consent, retention and deletion are workflow families—not policy documents disconnected from execution.
- 09
Country packs, never code forks
Tariffs, schemas, rules and integrations vary by pack so the product core remains one governed system.
The three stages of intelligence
Capability expands only when the control seam is ready.
The roadmap starts with bounded models inside deterministic workflows, then adds reasoning where unfamiliar exceptions justify it.
Stage 2
Consolidate
Heuristics mature into trained models on consented tenant data, with bounded single-shot LLM features.
- Freight audit
- Route and drayage optimisation
- Congestion forecasting and gap-filling
- Status summaries
- Query drafting
- Text-to-SQL analytics
Stage 3
Roadmap
Extend
Reasoning enters at exactly one seam: the exception queue.
- Bounded tools
- Read-mostly permissions
- Explicit step limits
- Human approval before irreversible action
Architecture trade-off
Why not agentic-first?
Agentic systems have real advantages. The question is where those advantages outweigh variability, cost and control requirements.
| Decision area | Deterministic core with embedded ML | Agentic-first |
|---|---|---|
| Time to first release | Wins · bounded workflows ship in testable increments | Slower · tool and guardrail surface must mature first |
| Testability | Wins · replayable states and contract tests | Probabilistic paths expand the test surface |
| Reliability under load | Wins · predictable execution and retry semantics | Long reasoning chains add latency and variance |
| Auditability for customs | Wins · named rules, versions and approvals | Requires additional reasoning traces and controls |
| Unit economics | Wins · models run only at defined decision points | Repeated reasoning raises variable cost |
| Change velocity | Wins · domain changes stay isolated and reviewable | Prompt and tool changes can interact unpredictably |
| Novel-exception handling | Bounded to known workflows and human queues | Wins · can compose tools for unfamiliar situations |
| Capability ceiling | Strong within designed process boundaries | Wins · broader reasoning once tools and controls are ready |
Agentic-first wins novel exceptions and the ultimate capability ceiling—which is exactly where we add it, when the seam is ready.
What we deliberately do not do
These are architectural constraints, not marketing qualifications. They keep authority, evidence and tenant boundaries explicit.
No model decides a statutory outcome.
No agent sits in the primary filing path.
No prompt-only business rules—rules live in decision tables domain experts own.
No cross-tenant data mixing without explicit consent.
Architecture questions
Control before cleverness.
How the execution model keeps probabilistic assistance inside auditable operational boundaries.