Architecture
The part that does not change when the vertical does.
Drevan is an argument that memory, orchestration, tool discovery, and provenance belong below the application, not inside it. This is how that argument is built.
The substrate
The defensible core, not another wrapper.
Anyone can call an LLM. The moat is everything around it — and it compounds with every agent, every call, every memory written.
MEMORY
Categorized memory (SGM)
Beyond vector RAG. Memories typed by role, trust, recency.
Recalled across calls in production — no DB query in the prompt.
ORCHESTRATION
Trust-aware orchestrator
Every action carries provenance. Zero-trust by construction.
LangGraph state machine with halt + escalation nodes — live.
MCP
MCP-native tool discovery
Agents grow their own capabilities. No hardcoded registries.
Tools register at runtime — no redeploy, no schema drift.
VISION
Multimodal grounding
MediaPipe + RT-DETR pipeline. Language tied to perception.
Scene state flows to the LLM as primitives, not raw bytes.
EVENTS
Event-backbone substrate
Async-first. Multi-agent coordination without polling.
Background lookups + analytics run off the hot path.
SECURITY
Zero-trust primitives
Built into the substrate, not bolted on top.
Guardrail-gated dispatch + approval queue for destructive ops.
Vendor-neutral by design — swap any layer
// Swap any layer — one env var change
// SUBSTRATE TOPOLOGY
provenance-tracked
Trust-Aware Orchestrator
categorized · typed
SGM Memory
tools grow themselves
MCP Discovery
MediaPipe · RT-DETR
Vision Pipeline
async · multi-agent
Event Backbone
policy at the substrate
Zero-Trust Guardrails
Decisions
Five choices that are load-bearing.
Each of these is reversible in principle and expensive in practice, which is what makes them worth writing down with the reasoning attached.
One config file selects every backend
Memory store, embedder, vision processor, event publisher, and orchestrator are chosen in a single config and instantiated lazily behind four abstract interfaces. Swapping a vector database is a config change, not a migration.
Why
Vendor lock-in in the AI stack is not hypothetical — providers change pricing, terms, and availability on short notice.
Verticals are configuration packs
Solar and real estate share one agent runtime and differ only in knowledge, persona, qualification logic, and compliance rules. A regression suite runs the packs on every change.
Why
If a second industry needs a fork, the substrate claim is false. This is the test that keeps it honest.
Two coordination topologies, deliberately
Agent teams run as parallel deliberation or as an ordered pipeline. Both bound their own control flow, so neither can loop indefinitely.
Why
Every additional topology multiplies the states the runtime has to be correct in — and unbounded agent loops fail as an invoice, not an error.
Security guards, not security conventions
Outbound requests, model-supplied regex, secret decryption, and inference spend each pass through one enforced guard, and each guard has a script that fails when a call site skips it.
Why
A convention is something people remember. Agentic systems execute far too fast for anyone to be remembering anything.
Migrate by wrapping, never by rewriting
New capability is added as an additive layer over seams that already take the relevant parameters, so the working path keeps working byte-for-byte while the new one is proven.
Why
A rewrite of a system that has live users is a bet that pays out only at the very end.
Honesty
What is not built yet.
Anyone evaluating this seriously will find the gaps, so they are stated here instead.
Scale is unproven
Everything here runs correctly at the volume it has seen, which is small. Nothing on this site should be read as a claim about behaviour under load.
One vertical is genuinely live
SuryaSetu carries real users. The second vertical pack exists and passes its regression suite, but has not been run against a real market.
No external users on the substrate
Every system on this site is built by the same person who built the substrate. That is the strongest possible test of coherence and the weakest possible test of developer experience.
Provenance is partial
Trust and provenance stamping is designed and partially implemented. It is on the roadmap, not in production, and is described that way everywhere on this site.
Why publish this