Applied Scientist — Generative AI & Multimodal Intelligence

Siddhant
Sanadhaya

I build agentic AI systems that reason, remember, and act — from multi-agent orchestration platforms to production computer-vision pipelines.

India · Open to relocation
View résumé Email LinkedIn

01 — Summary

Notebook to production, on repeat.

Six years of building software, five of them spent pushing generative AI, computer vision, and NLP systems out of research notebooks and into production. I design agentic orchestration platforms with LangGraph, ship AutoML pipelines that cut compute overhead, and build computer-vision systems that read handwriting, detect objects, and analyze medical images — deployed across AWS, Azure, and GCP for enterprise clients, with the applied-R&D instinct to turn a client's bottleneck into a working proof of concept.

6+ yrs
engineering experience5+ in GenAI, ML, CV & NLP
25%
efficiency gain from automationRoll Number Detection System
50%+
tech-writing workload cutGenAI Enterprise Writing Assistant
3 clouds
in productionAWS, Azure, GCP

02 — Skills Graph

The stack, mapped.

Drag a node. Hover to trace what connects to what — the tools, frameworks, and disciplines that actually overlap in the work.

03 — Method

How I build an agent.

Two phases. The first only adds. The second only decides.

Phase 01 · Integration · the object grows from the ground up

Five machines close on one agent build.

I treat a build like an assembly line — each stage adds exactly one governed capability to the same object, and nothing that goes on ever comes back off. Run it to the end and the agent is finished. Finished isn't the same as cleared to act.

ASM · BUILD SEQUENCE Scroll — you're driving the line
  1. 01Scope the decisionPick a decision worth automating and draw a hard boundary around what the agent may touch, before any model gets involved.
  2. 02Ground the contextWire retrieval to a real knowledge base with permissions and freshness, not a folder of PDFs.
  3. 03Connect the systemsGive it the smallest set of APIs and tools it actually needs, each one permissioned and reversible.
  4. 04Coordinate the reasoningRoute between models by cost and risk, and coordinate multiple agents with shared state and memory.
  5. 05Ship and watchCI/CD, canary rollout, and a feedback loop that pushes corrections back into the design, not a dashboard nobody reads.

What a stage can do

Add. That's the whole move — fit the part it was built for onto whatever's already standing, and the agent is more capable than it was a moment ago.

What no stage can do

Stop. Nothing in this phase outputs a no. Ask a build stage whether the agent it's assembling should be allowed to act, and it will just keep building.

Phase 02 · Governance · pre-flight, in-flight, after-launch

Finished isn't cleared.

Once the build is complete, it runs through a governance pass inspired by Microsoft's AI governance framework — a check before it ships, a check while it's live, and a check after it's in production.

PAD · FLIGHT SEQUENCE Autoplay · loops

Pre-flight

Four stations, bolted to the build

Agent Marketplace vets every part before it can enter a build. Agent Runtime caps the blast radius with no standing privilege. Agent Compliance lints the policy pack like code. Agent Control Specification gives every request the same verdict, replayable. Clear all four, or nothing lifts.

Mission control

Agent OS — the only station that can say go

The gate that turns four clearances into one launch call. It moves an agent through a lifecycle from draft to approved to retired, and nothing ships or acts without clearing it. Fail-closed: unsure means no.

In-flight → production

Four stations, armed at T+0

Agent Hypervisor traces every action back to the rule that allowed it. Agent Mesh names and scores every agent-to-agent call. Agent SRE holds it to an SLO with a stop button that works in one click. Agent Lightning lets it improve without drifting off policy. None of them switch off at the boundary — they ride straight through into production.

This build-and-governance framework is adapted from an interactive product experience I engineered at StatusNeo for Agent OS, an enterprise AI-governance platform.

04 — Execution Log

Six years, four roles, one thread.

Get applied AI out of the notebook and into something that runs.

  1. 2023 — present

    Software Development Engineer — AI & Generative Systems

    StatusNeo

    • Architected and deployed enterprise-grade Agentic AI Orchestrator solutions for procurement applications using LangGraph — multi-step reasoning agents, tool-based automated workflows.
    • Designed multi-user, multi-session state architecture with per-user chat persistence, real-time context management, and custom memory.
    • Engineered e-commerce recommendation systems ("Suggested Stores", "Buy Again") driving engagement and NPS.
    • Built large-scale semantic data pipelines with PySpark and Transformer models for text analytics across unstructured enterprise data.
    • Shipped end-to-end AutoML pipelines, cutting compute and operational overhead across training workflows.
  2. 2022 — 2023

    Python / Machine Learning Developer — Computer Vision & Cloud

    McKinley Rice Creativity

    • Led deep-learning and computer-vision research to improve accuracy and speed in medical image analysis, using PyTorch, OpenCV, and custom architectures.
    • Built high-throughput, low-latency RESTful microservices (FastAPI, Flask, Django) serving ML inference to production clients.
    • Designed serverless computer-vision pipelines on AWS Lambda for real-time image analysis at scale.
    • Partnered with product teams to turn client requirements into working proofs of concept.
  3. 2020 — 2022

    Project Engineer — Deep Learning & OCR Systems

    Rawattech

    • Led an offline, deep-learning multilingual handwritten OCR system — independently built the core parsing and character-recognition models.
    • Directed an automated Roll Number Detection System, lifting project efficiency by 25%.
    • Ran pre- and post-sales technical consulting, turning customer bottlenecks into AI-driven solution architectures.

05 — Key Innovations

Applied R&D that shipped.

Computer Vision · OCR

Multilingual Handwritten OCR & Roll Number Detection

Rawattech

An offline, deep-learning OCR system for multilingual handwriting — built the core parsing and character-recognition models from scratch — paired with an automated Roll Number Detection System for exam sheets.

25%higher project efficiency
Offlinefully on-device inference
Multilingualhandwritten character sets

Generative AI · Enterprise Tooling

GenAI Enterprise Technical Writing Assistant

StatusNeo

An LLM platform wired into Jira, Confluence, and Google Docs via automated CI/CD, generating technical documentation directly from the tools teams already work in.

>50%writer workload cut
3systems integrated
CI/CDautomated pipeline

06 — Education

Formal training, still compounding.