Proprietary AI Infrastructure
Not a chatbot. Not a wrapper.
A purpose-built intelligence
architecture.
Most organisations claiming AI are connecting a form to a general-purpose language model. Axon™ is different — a proprietary six-layer intelligence architecture designed for the evidence standards, accountability requirements, and execution complexity of global social impact work.
One intelligence layer. Eight products.
Every KairoPact product runs on the same Axon™ architecture — shared intelligence, isolated data, compounding learning.
Why general-purpose AI is structurally insufficient for this sector
Hallucination on specialist content
General-purpose language models produce plausible-sounding but factually unreliable outputs on development sector content. They systematically favour well-known organisations over high-impact field-stage implementers. They confuse funders, misattribute evidence, and generate submissions that read well but fail on compliance review. For work affecting capital allocation and programme outcomes, this is a disqualifying error pattern.
No domain memory or compounding intelligence
Consumer AI tools treat every session as new. No memory of past submissions. No accumulated understanding of funder preferences. No institutional knowledge of what interventions work in which geographies. A well-designed system knows more after a hundred engagements than after one. Axon™ is built on this principle. General-purpose tools are not.
No accountability layer
General AI produces outputs with no practitioner review, no audit trail, and no structured record of what judgment was applied and why. For recommendations affecting grant decisions, grantee selection, or capital allocation, this is not a design gap. It is a disqualifying condition.
"Axon™ was designed to address all three — not as features added later, but as design requirements from the start."
The Axon™ Lexicon
Axon™ uses a proprietary vocabulary that describes its architecture precisely. These are not marketing terms — each maps to a specific technical component.
A six-layer architecture. Each layer has a single responsibility. No layer may be bypassed.
Every client request flows down through six layers. Every output flows back up. This is not a single model handling everything.
Client Interfaces
Every human interaction with Axon™ originates here — practitioner dashboard, NGO portal, ImpactMatch consultant portal, API surface. Clients never interact with Axon™ directly. All outputs pass through a practitioner before reaching a client.
Gateway & Session
Authentication, request routing, and rate management. Full Stream checkpoint state is persisted at every Practitioner Gate — so a practitioner in Nairobi can review a Gate output at 9pm and another in London can resume the same Stream the following morning with complete context and zero risk of duplicate outputs.
Stream Orchestration
The workflow logic layer. Each Axon Stream™ runs here — sequenced Axon Nodes™, conditional routing, shared state, and Practitioner Gate checkpoints. Axon™ uses a framework-agnostic orchestration approach: all orchestration logic maps across frameworks, so the underlying runtime can scale without touching the data architecture, memory model, or Gate pattern. Built for 50+ concurrent Streams across multiple time zones.
Multi-Model Routing
Every language model call from every Axon Node™ routes through a central routing layer. No Node calls a model provider directly — this is enforced at code review. The router handles model selection, cost tracking in real currency per Node per run, retry logic with hard stops, and fallback. The architecture is model-agnostic: the underlying models can be upgraded without changing any Node logic.
Axon Spine™ (Storage & Knowledge)
Three stores working in concert: PG Vector (semantic retrieval, namespace-isolated per client, VectorStore abstraction makes the backend swappable), Supabase PostgreSQL (structured system of record — every Node output, every Gate decision, every cost log), and the Feedback Store ({prefix}_decisions — immutable practitioner judgment data, the compounding intelligence moat).
Observability & Operations
Distributed tracing across every Node call and every model invocation. Per-Node cost tracking. Stuck-Stream detection every 5 minutes. GDPR / India DPDP Act / UAE PDPL / Singapore PDPA compliant audit logs built into every Stream run from the start — not retrofitted. Every Vault access and every Gate decision is logged with timestamp and actor identifier.
Different tasks require different models. Axon™ routes intelligently between them.
A single model handling all Nodes produces homogenised outputs — submissions that sound identical, analysis that defaults to the same framing, matching that reflects one model's training biases.
Node diversity prevents homogenisation
When different Nodes in a Stream use different models, outputs genuinely differ — which is a precondition for the Critic Node architecture. A Critic Node checking one model's synthesis with a different model's assessment produces a fundamentally stronger quality signal than the same model reviewing its own output.
The synthesis Node anchors quality
Regardless of which models upstream Nodes used — fast extraction Nodes, ideation Nodes, bulk processing Nodes — the synthesis Node that assembles final client-facing output always runs on the highest-quality available model. Speed is optimised where it matters. Quality is non-negotiable where it matters.
Hard stops are structural safeguards
Axon Streams™ without cost ceilings would run indefinitely on failed reasoning loops. Every Node has a maximum attempt count and a token hard stop. If quality gates are not met within the attempt budget, the Stream terminates with a logged error rather than looping silently.
Intelligence across formats — not just text.
The development sector does not communicate in text alone. Field evidence arrives as photographs. Training materials are video. Community interviews are audio. Survey instruments produce structured data. An intelligence architecture that processes only text is, by definition, incomplete.
Axon™ uses a natively multimodal embedding model that represents and retrieves knowledge across formats — covering 100+ languages used across South Asia, MENA, Sub-Saharan Africa, and Asia Pacific.
Documents & Text
Grant proposals, field reports, research papers, evaluation documents, policy briefs, donor correspondence. Embedded with multilingual support — so a funder intelligence query in English can surface evidence documented in Hindi, Swahili, Arabic, or French. Cross-lingual retrieval is built in, not bolted on.
Audio & Voice
Field interview recordings, community voice notes, WhatsApp audio messages from programme staff, stakeholder call summaries. Audio is embedded alongside its text transcript — making what was said in a field call three months ago retrievable as context for the next programme review.
Video & Visual
Training materials, field documentation photographs, community mapping outputs, video field reports. Visual content is processed and embedded alongside text metadata — so photographic field evidence and written field reports are retrievable in the same semantic query.
Structured Data
Survey results, monitoring indicators, expenditure data, beneficiary counts, baseline and endline figures. Structured data is embedded in formats that enable semantic retrieval — so a question about programme performance can surface both a written evaluation and the underlying dataset that informed it.
"Development sector intelligence does not live in documents alone. It lives in what a field officer said on a call, what a photograph of a construction site shows, and what a dataset reveals when compared to a baseline taken two years earlier."
A three-tier memory architecture. Intelligence that compounds with every engagement.
Most AI systems are stateless — each request independent, no memory of what came before. Axon™ is built around a fundamentally different model: three distinct memory tiers, each with defined responsibilities, retention rules, and retrieval logic.
Tier 1
Working Memory
Scope: Active Stream run only
What it holds: Full Stream checkpoint state at every Gate boundary. The practitioner dashboard snapshot. Run-level context that Nodes share as a Stream executes.
Key rule: Working Memory never crosses Stream boundaries. It is cleared after run completion.
Why it matters: A Stream paused at a Practitioner Gate for 12 hours can resume with complete context — every Node output, every intermediate result, every decision made so far — exactly where it stopped.
Tier 2
Episodic Memory
Scope: Indefinite — immutable, never deleted
What it holds: Every Practitioner Gate decision — the original Node output, what the practitioner changed, and why (written rationale + word-level diff). Every Critic Node score. Every cost log.
Key rule: Append-only. No update, no delete. The record is permanent.
Why it matters: This is the most strategically valuable data Axon™ produces. Encoded expert reasoning about specific decisions in specific contexts — at scale, over time. It is the primary input to Axon Loop™ and the foundation of the compounding intelligence moat.
Tier 3
Long-Term Memory (LTM)
Scope: Subject to TTL policy and importance scoring
What it holds: Client documents, winning proposal patterns, sector evidence studies, funder intelligence, field data. Stored in namespace-isolated Vaults per client. A shared sector vault holds PII-stripped, human-approved knowledge accessible across all Streams.
Key rule: Retrieval is governed by a composite re-ranking formula — not raw cosine similarity alone. Recency, access frequency, and named entity match all contribute to which knowledge surfaces.
Why it matters: A client who has run twenty grant qualifications through Axon™ benefits from twenty cycles of funder intelligence and institutional memory applied to every new brief. The system knows more. Not because it was retrained — because it remembered.
The system challenges its own outputs. Then improves from the result.
A single Node generating a single output has no mechanism to catch its own errors. Axon™ embeds Critic Nodes — independent intelligence units that review the outputs of generating Nodes before those outputs proceed in the Stream.
Critic Nodes
How it works
The Critic Node runs in parallel with the primary generating Node. It reviews the output for logical inconsistencies, unsupported claims, compliance gaps, and evidence misattributions — producing a structured critique that surfaces alongside the primary output at the Practitioner Gate. The practitioner sees both the output and the system's own assessment of its weaknesses.
Why it matters
In a multi-Node Stream, errors at early Nodes cascade. A funder intelligence Node that misclassifies a requirement will corrupt every downstream Node that depends on that output. The Critic Node breaks this cascade at the source — not seven Nodes later when the damage is compounded and the Gate review has already been completed.
Axon Loop™
Three feedback loops that run continuously:
Loop 0 — Skills to Nodes
Validated outputs from real engagements graduate into production Node instructions. No work is discarded. Every real engagement improves the starting quality of the next one.
Loop 1 — Node Improvement
A meta-Node reviews Episodic Memory patterns after every 10 approved runs per Node — identifying recurring failure modes and proposing Node instruction updates. The founder approves every proposed change before it goes live. No Node updates itself autonomously. Ever.
Loop 2 — Output Quality Tracking
Every human edit at a Practitioner Gate is measured as a percentage of the Node's output that was changed. Nodes below 30% edit rate are at production quality. Nodes above 70% are flagged for Loop 1 review.
"No competitor can buy this data. No competitor can replicate it without running the same volume of real engagements with the same quality of human oversight."
Every consequential output passes a Practitioner Gate.
The Practitioner Gate is not a quality check bolted on at the end. In every Axon Stream™, it is structural — mandatory checkpoints at each phase boundary, requiring practitioner review before the Stream proceeds.
Approve
The practitioner reviews the Node output and confirms it meets quality and accuracy standards. A minimum rationale is required — even for approvals. The Stream resumes from the next Node.
Override with Rationale
The practitioner modifies the output and provides a written reason — a specific correction, an evidence note, a judgment call. The Stream resumes with the corrected output. The original Node output is never modified — it is frozen. The override is a new record, with a structured diff capturing exactly what changed.
Escalate
The Stream is paused indefinitely for senior review. Used when an output raises a concern beyond the reviewing practitioner's remit. The Stream state is fully preserved. Nothing is lost.
The rationale written at every Gate — what was changed, why, and in what context — is written as an immutable record to Episodic Memory. It cannot be edited, deleted, or overwritten. It is the compounding intelligence moat.
Client data is isolated, encrypted, and never used to improve outputs for other clients.
Vault Isolation
Every client has a dedicated Axon Vault™. Private vector namespace. Private artifact store. No Node working for one client can retrieve from another client's Vault. Enforced at query construction and at the database row-level security layer. Two enforcement mechanisms, not one.
No Cross-Client Training
KairoPact does not use client data to train or fine-tune any model. Client data in the Axon Vault™ is used only for that client's Streams. It is never aggregated, never shared, and never used to improve outputs for other clients.
Full Audit Trail
Every Vault access, every Node retrieval from the knowledge Spine, and every Practitioner Gate decision is logged with a timestamp and actor identifier in a compliance-queryable audit log. Built into every Stream run from the start. Not retrofitted.
Regulatory Compliance
The architecture is built to satisfy GDPR, India's Digital Personal Data Protection Act 2023, UAE Personal Data Protection Law, and Singapore's Personal Data Protection Act — from the first client run. Not as an afterthought.
Independently reviewed. Not self-assessed.
The Axon™ architecture has been reviewed by four independent experts with AI infrastructure experience across healthcare, financial services, and the development sector.
Head of AI Engineering, global enterprise SaaS platform (10,000+ enterprise clients)
Validated multi-agent pipeline architecture with stateful orchestration as appropriate for HITL-heavy production workflows at scale. Confirmed multi-model routing as architecturally correct and hallucination-mitigating.
March 2026
Founder and former Field CTO, AI-native workflow products for financial services clients across ASEAN and Japan
Contributed the diagnostic intake architecture concept. Confirmed workflow families operate as a connected value chain. Validated intelligence layer as the primary competitive moat.
March 2026
Senior Solution Architect, global IT services firm. 15 years AI/ML. Healthcare AI implementations for major US insurers and hospital networks.
Validated stateful multi-agent HITL pipeline design. Confirmed Critic Node pattern as robustness mechanism. Confirmed hard stops and token limits as essential production safeguards.
February 2026
Two independent AI architects conducting Series A due diligence assessment
Both independently identified the three-table artifact store schema as the strongest architectural decision — specifically because most AI pipeline teams discover the need for structured run history too late. Axon™ made this decision at design time, before the first client run.
March 2026
Built for the rigour this sector demands.
The development sector moves capital that affects millions of lives. The AI infrastructure behind that work should meet a correspondingly high standard — explainable, auditable, practitioner-governed, and designed with the specific evidence standards and compliance requirements of global social impact work.
Axon™ is KairoPact's investment in that standard. It is not finished. It is being built with discipline, reviewed externally, and improved through every engagement. The architecture is documented. The build sequence is defined. That transparency is deliberate. It is the only thing that distinguishes a serious AI infrastructure investment from a marketing claim.