AI ENGINEER · AGENTIC SYSTEMS

Jeet Soni

I build AI agents that survive production, and the platforms they run on.

GitHub
LinkedIn
YouTube

5+

Years of engineering

10+

Production AI agents shipped

17

LLM agents in one pipeline

6

Industries served in production

AI Engineer with 5+ years across full-stack and applied-AI development, now leading both the build of an enterprise LLM agent platform and its delivery to clients who bet real workflows on it.

As founding engineer of AgentOS at AvestaLabs, I shaped an agent orchestration platform from whiteboard to production: context engineering, RAG pipelines, vector-backed agent memory, MCP client and server, custom tools, guardrails.

I sit on both sides of the table: architecting the platform, then embedding with enterprise clients to map their workflows and ship agents that hold up when real users arrive. Ten-plus production GenAI agents across finance, legal, media, real estate, e-commerce and travel.

Everything runs on a full-stack spine of Next.js, Node.js, TypeScript and Python, with an architect's bias for ports & adapters, DDD, and systems that can swap their LLM without rewriting their soul.

Demos are easy.
Production is the product.

Working philosophy

Trajectory

Dec 2024 → Present

AvestaLabs

AI Product Engineer & Team Lead

Avesta HQ's AI division

  • Founding engineer of AgentOS, an enterprise suite to design, deploy, evaluate and monitor AI agents; lead the core engineering team.
  • Architected Efficia, the agent orchestration product: context engineering, RAG, vector-backed agent memory, MCP client/server, custom tools framework, guardrails.
  • Embedded technical lead on enterprise engagements, with 10+ production agents live across finance, legal, media, real estate, e-commerce and travel.
  • Shipped OnlyFacts: climate questions answered as real-time generative UI. India Avenue: a customer-facing fund-data agent. Legal automation: lawyer-ready contracts drafted in 3-5 minutes.
AgentOSEfficiaMCPRAGGuardrailsTeam Lead

May 2023 → Dec 2024

Avesta HQ

Software Engineer

Ahmedabad, India

  • Spearheaded the complete rebranding and feature expansion of view.com.au, a major Australian real-estate portal, on Next.js, Node.js and PostgreSQL.
  • Led end-to-end delivery of product features, working directly with PMs and clients on requirements.
  • Designed implementation architecture and mentored the team on Next.js best practices and design patterns.
Next.jsNode.jsPostgreSQLArchitecture

Jan 2021 → May 2023

Space-O Technologies

MERN Stack Developer

Ahmedabad, India

  • Built and maintained MERN applications: analytical dashboards, CRM web apps, delivery platforms.
  • Integrated secure payment gateways (Stripe, TYRO) across multiple web applications.
  • Star Performer of the Month.
ReactNode.jsMongoDBStripe

Dec 2019 → Sep 2020

9Brainz · TRUESYS

MEAN / Web Developer

Early career

  • Engineered MEAN-stack web APIs and real-time Angular applications; AWS API Gateway, Lambda, DynamoDB for cloud-native projects, growing from intern to live-project delivery.
AngularAWSNode.js

Selected Builds

Swipe the cards →

LIVE
01

KalpanaAI

AI Video Production System · Founder

A prompt-to-video product where AI plans the story, generates executable motion graphics, and hands a reviewed composition to a durable rendering workflow.

  • Clean Architecture across domain, application, infrastructure and presentation layers, with PipelineJob owning legal state transitions and artifact rules
  • BullMQ workers execute persisted stages with explicit retry policies; PostgreSQL checkpoints scripts, fact checks, audio, directions, code and rendered assets
  • AI-generated React/Remotion scenes are compile-checked, inspected for runtime hazards and repaired before composition; MCP client scene submissions also pass bounded three-frame smoke renders
  • Redis pub/sub streams live progress while buffered scene events support replay after a client reconnects
TypeScriptNext.js 15Express 5BullMQRemotionPostgreSQLCloudflare R2
Read case study→
Open live product↗
KalpanaAI studio: turn any idea into a finished video
LIVE
02

DICOM Viewer + PACS

Scanner-to-Report Radiology Platform

A complete imaging chain built solo: scanners push to an offline-first edge gateway, studies land in a multi-tenant cloud PACS, and an AI agent reads scans by driving the viewer itself.

  • Pure-DICOMweb archive discipline (STOW-RS, QIDO-RS, WADO-RS): Orthanc swappable for AWS, Google or Azure health imaging by config; the archive is never exposed to the browser
  • Edge gateway for clinics: scanners C-STORE to a LAN node; a standard-library-only Python forwarder pushes stable studies to the cloud and retries forever through outages
  • Agentic deep read with the viewer as the AI's hands: view_slices and measure_hu execute in-browser on an off-screen viewport; models scored against a pydicom ground-truth harness
  • MPR crosshairs, 3D volume rendering and quad layouts that survive older iPhones: live WebGL2 probing sizes each volume to the device's real GPU ceiling
Next.js 16React 19Cornerstone3DPostgreSQLOrthancVercel AI SDK
Read case study→
Open live demo↗
DICOM viewer 3D workspace: MPR slices and volume-rendered skull of a head CT with signed report
LIVE
03

Redline Agent

Rulebook-Driven Contract Review · Open Source

Upload a contract and a negotiation playbook: an AI agent redlines it clause by clause and exports native Word tracked changes, with grounding evals that prove every edit cites a real rule.

  • Real OOXML revisions (w:ins/w:del): the exported .docx opens in Word with working Accept/Reject
  • “No citation, no redline”: deterministic grounding checks plus run-level evals for citation coverage and anchoring
  • Clause-by-clause SSE streaming with reconnect replay; gap analysis drafts ready-to-insert missing provisions
  • Deliberately MERN: MongoDB Atlas, Express, React 19, Node; model-agnostic via Vercel AI SDK + AI Gateway
MongoDB AtlasExpressReact 19TypeScriptVercel AI SDKDocker
Open live demo↗
Redline Agent review board: clause list with risk badges and a tracked-changes diff citing playbook rules
IN PRODUCTION
04

AgentOS · Efficia

Enterprise Agent Platform · AvestaLabs

The factory that builds the agents. An enterprise suite to design, deploy, evaluate and monitor AI agents, architected from concept to production as founding engineer.

  • Agent orchestration with context engineering, RAG pipelines and vector-DB-backed agent memory
  • MCP client & server, a custom tools framework, and guardrails as first-class citizens
  • Eval-first architecture: simulate messy conversations, evaluate every edge case, auto-correct and loop
  • Companion products for LLM observability, RAG data-ingestion and evaluation
TypeScriptNode.jspgvectorMCPLangfuseEvals
Avesta AgentOS: AI co-worker teams running sales, finance and support workflows in parallel
SHIPPED
05

view.com.au

Australian Real-Estate Portal

Complete rebranding and feature expansion of one of Australia's major property portals, led end-to-end on Next.js, Node.js and PostgreSQL.

  • Led feature delivery working directly with PMs and clients
  • Designed the implementation architecture and mentored the team on Next.js patterns
Next.jsNode.jsPostgreSQL
view.com.au↗
view.com.au homepage: property search Australia-wide

More on
GitHub

github.com/jeetsoni ↗

One concept. One animation. One minute.

SynapByte is my animated engineering channel: one concept, one animation, one minute. Every single video is generated end-to-end by KalpanaAI, the platform I built. The content is the demo.

SynapByte · Broadcast

On Air

Nginx Architecture: How One Process Handles 10,000 Connections

Deep dive · JUL 11

YouTube ↗

Cross-posted as reels on Instagram

@synap_byte ↗@jeetsoni.ai ↗Subscribe ↗

Most-watched · @jeetsoni.ai reels

Instagram

1,400+

followers

@synap_byte ↗

AI & software engineering, animated. LLMs · system design · dev internals.

YouTube

73+

videos

SynapByte ↗

Nginx internals, UUID vs integer keys, CI/CD from scratch. Where every concept clicks.

Instagram

4,800+

followers

@jeetsoni.ai ↗

JS → production AI systems. RAG · agents · tools · UI. Real-world AI, not demos.

Arsenal

AI / LLM

AI Agents & Multi-Agent OrchestrationRAG PipelinesMCP (Model Context Protocol)Prompt & Context EngineeringFunction / Tool CallingEmbeddings & Vector Search (pgvector)Agent Evaluation & GuardrailsLLM Observability (Langfuse)Claude · Gemini · OpenAIVercel AI SDKMultimodal / VisionStreaming

Languages & Frameworks

TypeScriptJavaScriptPythonNext.jsReactNode.jsExpressNestJSAngular

Architecture

Clean / Hexagonal ArchitectureDomain-Driven DesignSOLIDDesign PatternsProperty-Based TestingCI/CD

Data · Cloud · DevOps

PostgreSQL + pgvectorMongoDBRedis · BullMQPrismaGraphQLAWS Lambda · API Gateway · DynamoDBCloudflare R2DockerTurborepoServerless

Let's ship
something real.

Building an agent platform, wiring AI into a product, or hunting for an engineer who treats evals as seriously as demos? My inbox is open.

adesharajeet@gmail.com