0%
SPSathwik PrathiLead AI Engineer

Lead AI Engineer · Circuitry.ai

I build systems thatthink at scale.

Sathwik Prathi — I build AI agents that make decisions, not just conversation: claim scoring, autonomous decisioning, and the reliability layer that makes it safe to let an agent move money.

Scroll
Agentic AIMulti-Agent OrchestrationA2AMCPAutonomous DecisioningFastAPIKubernetesLLM SystemsAgentic AIMulti-Agent OrchestrationA2AMCPAutonomous DecisioningFastAPIKubernetesLLM Systems

— / What it looks like

A conversation with the systems I build

agent-session · claims-platform
you

Why was claim #48213 auto-approved instead of escalated?

Scoring claim against coverage policy…
Calling MCP tool → get_claim_history(claim_id)
Checking anomaly-detection thresholds…
agent

Trace:coverage_check ✓anomaly_check ✓payment.release()

01 / Selected Work

Five problems worth telling you about

01 / Lead

Agent-to-Agent Claim Decisioning

Replaced a monolithic single-agent flow with specialized agents for intake, document understanding, coverage validation and anomaly detection — coordinated through supervisor-pattern handoffs and MCP tool calling.

Agentic AIA2AMCP
02 / Lead

The Reliability Layer

Durable state checkpointing so long-running workflows survive crashes and redeploys, idempotency keys on every side effect, and compensating actions for partial failure — the plumbing that makes it safe for an agent to move money.

Distributed SystemsKubernetesReliability
03 / Senior

Agentic RAG Planning

A planning agent that decomposes complex queries into parallel RAG workflows with reranking — positive ratings on multi-step questions went from 95% to 99%.

RAGRerankingAgentic AI
04 / SE

Dialogue Intelligence Layer

Replaced a legacy Rasa-based system with a proprietary FastAPI dialogue layer, plus a vision-based chunking pipeline that recovered tables and diagrams OCR was silently dropping — cutting costs 60–70% and pushing document coverage past 99%.

FastAPILLMComputer Vision
05 / Personal

Eye Wellness Tracker

Real-time blink detection desktop app (MediaPipe, PyQt6) with an offline-first telemetry backend — SQLite outbox, async FastAPI, idempotent ingestion for unreliable networks.

MediaPipePyQt6FastAPI
Drag or scroll sideways

02 / About

I build AI agentsthat make decisionsnot justconversation.

At Circuitry.ai I lead the agents team behind our decision intelligence platform — agent-to-agent orchestration that scores a claim, then decides what happens next: release payment, retry a failed downstream call, or escalate to a human adjuster with the full reasoning trace attached.

0%cut in dialogue system costs
0%enterprise document coverage
0promotions in 16 months

03 / Experience

Where I’ve been

Click a role to see what actually shipped.

2017 — 2021B.Tech, Electrical Engineering IIT Dharwad · GPA 7.9/10

— / Stack

Toolbox

ai & agentsLLMs ·RAG ·Agentic AI ·MCP ·A2A ·LangGraph ·LangChain ·CrewAI ·OpenAI API ·Anthropic API
backendFastAPI ·Django ·DRF ·Flask ·SSE streaming ·JWT
ml & cvPyTorch ·Scikit-learn ·BERT ·OpenCV ·MediaPipe
infraAWS ·GCP ·Docker ·Kubernetes ·Temporal ·Redis ·OpenTelemetry ·Datadog
dataPostgreSQL ·SQLite ·Milvus ·NoSQL
currentlytightening agent reliability evals,rebuilding this site

04 / Writing

Notes from the work

I'm starting to write about agent-to-agent orchestration, MCP, and the reliability problems nobody mentions in the demo — the part between "it works" and "it's safe to let it move money."

Read on Medium →

05 / Contact

Let's build agents
worth trusting.

Say hello →
sathwikprathi7@gmail.com