01 / SUBSTRATE

The substrate for agentic AI.

Every AI agent needs memory, orchestration, and tool discovery. Today every team rebuilds them — badly. Drevan ships them as infrastructure.

// Beyond the wrapper. Beneath the agent.

02 / THE PROBLEM

Every AI startup is solving the same 5% of the problem.

The prompt layer is solved — and commoditized. The 95% that isn't: memory that survives sessions, orchestration that respects trust boundaries, tool discovery that doesn't hardcode, grounding that doesn't hallucinate. Every agent company is about to hit this wall at the same time. We build the layer they'll reach for.

// Vendor-neutral by design — your stack survives the next model release.

03 / ARCHITECTURE

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

Model
Memory
Voice

// 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

04 / VERTICAL PROOF

SuryaSetu: the substrate, proven on real phone calls.

Solar lead-gen in India — regulated, Hindi-first, voice-only. The hardest possible test for an autonomous agent. Built by one person on Drevan, and now making real cold calls to real homeowners.

First live pilot · 2026.06.15

9

Real outbound calls

~70%

Agreed to a site visit

100%

Autonomous — no human on the line

n=9

Single dial-day · directional

// Directional, not yet statistically meaningful. We're confirming the consent rate across more dial-days before scaling ad spend — rate over count.

Live pipeline — SuryaSetu

INPUT

Voice call

VAD

Silero

STT

Sarvam

SUBSTRATE

Drevan

LLM

Groq

TTS

Bulbul

TRANSPORT

LiveKit

Barge-in detection · SGM memory across calls · trust-gated qualification decisions

LIVE IN DREVAN

Not a roadmap. Working software.

Every capability below is built and wired into the substrate today. The voice and calling stack is proven on real outbound calls; the rest is integrated and running in dev — honestly labelled, no “coming soon.”

Voice Pipeline

LIVE — PROVEN ON REAL CALLS

Real-time voice agent: Silero VAD barge-in detection → Sarvam STT (Hindi/English) → Groq LLM → Bulbul TTS → LiveKit WebRTC. Sub-2s round-trip in field conditions.

Inbound & Outbound Calls

LIVE — PROVEN ON REAL CALLS

Agents make and receive phone calls via LiveKit WebRTC. SuryaSetu cold-calls leads, qualifies them, and logs decisions to SGM — fully autonomous, no human on the line.

SGM Memory — Cross-Call Recall

BUILT

Semantic Graph Memory persists intent, outcomes, and consent across every call. A lead who mentioned "roof space" in call 1 is remembered in call 3 — without a database query in the prompt.

LangGraph Orchestration

BUILT

State-machine agent flow via LangGraph. Deterministic qualification branches, escalation nodes, and halt conditions — not a prompt loop hoping the LLM decides correctly.

MCP Runtime Discovery

INTEGRATED

Tools register at runtime via Model Context Protocol. New capabilities appear without redeployment — no hardcoded tool lists, no schema-drift failures.

Vision Pipeline

INTEGRATED

MediaPipe + RT-DETR wired into agent context. Visual state (objects, faces, scene) flows into the LLM as structured primitives — not raw image bytes.

Screen Automation

BUILT

Computer-vision-based UI control: EasyOCR + YOLO + template matching. Agents click, type, and verify at pixel accuracy — DPI-aware, with human-like mouse paths and retry logic.

Device Task Engine

BUILT

Windows system automation: open apps, move/copy/delete files, search directories, read system state. Path-restricted, extension-whitelisted, with an async approval queue for destructive operations.

WhatsApp & Email

INTEGRATED

Agents send WhatsApp messages and emails via fuzzy contact resolution — no exact name required. Confidence-scored matching before dispatch, guardrail-gated by default.

Google Calendar

INTEGRATED

Create, list, check availability, daily briefing, delete events via OAuth2. Natural-language scheduling wired into agent context — "find me a free slot Thursday afternoon" works.

Proactive Agent

BUILT

Unsolicited execution without a user prompt — scheduled briefings, weather checks, delegated task monitoring. Cron-like scheduler with a daemon runtime and event-log replay.

05 / VISION

One vertical today. The agent substrate tomorrow.

We earn the right to be the layer beneath every agent by first shipping a vertical that works in the hardest conditions — then opening the substrate beneath it. Wedge first. Platform next.

NOW

One vertical, in market

SuryaSetu — an autonomous voice agent making real calls in a regulated, Hindi-first market. Proof the substrate survives the real world.

NEXT

The substrate, opened

The memory, orchestration, and trust layer underneath SuryaSetu — packaged for the agent teams who need it and don’t want to rebuild it.

THE BET

The default agent infrastructure

Every agent company hits the same wall — memory that forgets, orchestration that drifts, tools that hardcode. Drevan becomes the layer they reach for.

// The wrapper era is ending. The infrastructure era is starting.

06 / ENGINEERING LOG

Open notebook. Closed source.

Every major architectural choice — documented, dated, defended.

First live pilot: 9 autonomous cold calls, ~70% agreed to a site visit

SuryaSetu made its first real outbound calls — fully autonomous, Hindi-first, founder-verified on tape. What broke, what held, and why we optimize consent rate before volume.

Voice pipeline on Drevan: Silero VAD → Sarvam STT → Groq → Bulbul TTS

End-to-end voice agent on the substrate. Latency profile, barge-in detection, LiveKit WebRTC — and why we chose Sarvam over Deepgram for Hindi.

LangGraph + MCP phase 1: ComponentFactory wired to runtime tool discovery

Dynamic tool registration at runtime — no hardcoded imports. The architecture decision that will matter when you add your 12th tool.

SGM schema, category taxonomy, and the consent model — design notes

How we designed the Semantic Graph Memory layer: write guards, cross-session retrieval, and why flat vector DBs fail at agent memory.

// New entry roughly every two weeks.

07 / EARLY ACCESS

I talk to three builders a month.

Early access is not a waitlist.

I onboard three builders a month. You bring a real agent product, a real production problem, and 30 minutes a week. I bring the substrate and ship to your specific constraints.

No demo calls. No sales deck. If you're building with agents and hitting the substrate problems — memory that breaks, orchestration that drifts, tool discovery that's hardcoded — write directly.

AB

Abhitesh Bhardwaj

FOUNDER · DREVAN

Ships production agentic systems at a venture-backed AI company. Drevan is the substrate he wished existed — so he's building it, and proving it on a live revenue vertical (SuryaSetu) at the same time.

Write to

abhitesh@drevan.io

Include: what you're building · where orchestration/memory is breaking · stack

We don't do RAG-as-a-service. We don't do chatbots.