Skip to content

Running in production

The substrate
beneath the agent.

Every agent needs memory, orchestration, and tool discovery. Today every team rebuilds them — badly, and from scratch. Drevan ships them as infrastructure, then proves they hold on a live autonomous voice vertical carrying real users under Indian data regulation.

Beyond the wrapper. Beneath the agent.

Recorded call — SuryaSetu, June 2026Awaiting consent clearance before publication

A real call from the pilot will sit here: the agent qualifying a homeowner in Hindi and English and booking a site visit, with the transcript following the audio. It is not published until the recording is cleared and identifying details are removed.

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.

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

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

What is built

Four systems. One claim.

Each of these runs on the same substrate, and each exists to prove a different property of it. None of them is a side project.

It survives real users and real regulation.

An autonomous voice agent that calls consented homeowners in Hindi and English, qualifies them, books site visits, and hands them to installers verified against a government empanelment list. Live in production.

~70% interest rate on 9 dialled leads

suryasetu.online

It coordinates many agents without becoming spaghetti.

A multi-tenant platform where agent teams run as either parallel deliberation or an ordered pipeline. Two topologies, deliberately not more, with security invariants enforced by scripts rather than convention.

~83 API handlers, 5 enforced invariants

getsios.xyz

It holds under real-time constraints.

A sub-second speech loop built for Indian languages and mid-sentence code-switching, with barge-in, false-interruption recovery, and per-stage latency decomposition in every call report.

Latency decomposed per turn, per stage

It runs unattended and measures its own output.

Six content pipelines sharing one config and one knowledge system, with a measurement loop that compares posts at the same age rather than on lifetime totals — and found a 125x spread between the best post and the worst.

95.9K views in 30 days, from 156 followers

@gattu_ai__

Engineering log

Open notebook. Closed source.

What shipped, when, and what broke on the way. The failures are here because a log that only records wins is marketing, not evidence.

  1. SuryaSetu

    Bounded reconcile to a time window — it resurrected deleted data

    The tool built to recover leads the Demand Engine never delivered had no time bound, so re-running it restored records that had been deliberately removed. Recovery tools need a window, or they are an undelete nobody asked for.

  2. SuryaSetu

    Wrote up the four silent failures of the first campaign day

    Every one of them reported success while doing the wrong thing. The pattern is worth more than the fixes: a pipeline that cannot fail loudly will fail quietly instead.

  3. SuryaSetu

    Fetched lead answers from the Graph API — every lead was arriving empty

    The answers were never in the webhook payload. The ingest path had been reporting healthy while writing blank rows for an entire campaign day.

  4. SuryaSetu

    Stopped hiding broken leads

    Malformed leads were silently skipped so the dashboard stayed clean. Now they surface as errors. A dashboard that is quietly wrong is worse than one that looks bad.

  5. Infrastructure

    Health watchdog, and the real shape of the idle-reclamation risk

    Free-tier compute can be reclaimed when it looks idle. The watchdog exists because the failure mode is not a crash — it is a machine that quietly stops being yours.

Read the full log →

The ask

Not raising today. Building toward a pre-seed round in Q1 2027.

If you invest in AI infrastructure, the engineering log is a better use of your time than a cold meeting. Come back when the substrate has more verticals on it.