AI-First Backend Engineer · M.Tech, IIT Hyderabad

Papagari
Sumanth Reddy

I build production RAG pipelines, agentic AI systems, and optimization engines that ship — from a solver that cuts energy-market runtimes by up to 71%, to an edge LLM that runs fully offline on a child's phone in under 450 ms.

Top 5%
GATE 2024 CS/IT
of 123,000 candidates
8.73
CGPA
M.Tech, IIT Hyderabad
71.3%
Best-case solver
runtime cut
<450ms
Edge-AI response
latency, fully offline
Portrait of Papagari Sumanth Reddy

Warm-start vs. cold-start solves

Real benchmark results — Day-Ahead Market MIP solver, Greenko Group internship, 2026

Test 7 49.9%
Test 8 67.2%
Test 1 70.9%
Test 3 71.3%

% reduction in solver runtime vs. a cold start, across 8 benchmarked test cases

About

Systems engineering, applied to AI that has to hold up

Sumanth is an AI-first backend engineer and recent M.Tech graduate in Techno-Entrepreneurship from IIT Hyderabad, where he placed in the top 5% of the GATE 2024 CS/IT exam nationwide. His work sits at the intersection of systems engineering and applied AI — scalable microservices, agentic LLM workflows, and production RAG pipelines that hold up under real operational constraints: market gate-closures, offline devices, ISO-compliant documentation.

Most recently, he spent five months inside Greenko Group's Gen AI / Operations Research team, shipping a solver-acceleration proof of concept for grid-scale energy trading and a production AI agent for enterprise documentation — while finishing an edge-AI thesis on the side. Along the way, he's worked with student founders and researchers in Nagoya, Kyoto, and Shimane, and coordinated hospitality for SPIC MACAY's 10th International Convention at IIT Hyderabad.

He's especially drawn to problems where reliability matters as much as capability: a warm start that fails safely, an LLM that stays grounded in its sources, a pipeline built to survive production traffic rather than just a demo.

Based in
Hyderabad, India
Currently
M.Tech graduate, open to AI/backend roles
Focus areas
RAG systems · agentic AI · backend infra · optimization

Featured build

GIMS AI Agent — automating ISO documentation at Greenko Group

The first in-house tool to generate native, editable IMS swimlane flowcharts — directly inside procedure documents.

"Automate procedure restructuring and flowchart generation while preserving full editability in Microsoft Word — without outsourcing."

Situation

Greenko's IMS procedure documents follow a strict ISO template — swimlane flowcharts, standards references, KPI tables, document control metadata. Every update was rebuilt by hand: slow, and easy to get wrong.

Action

  • Split the work across 5 specialized pipelines — full restructuring on AWS Bedrock Claude, native Word DrawingML flowchart generation, prompt-driven document generation over SSE, an agentic chat assistant on self-hosted vLLM Qwen with web search for standards, and dual-mode template/input generation
  • Shipped a Dockerized React + TypeScript multi-session interface with streaming responses, drag-and-drop uploads, and job-status tracking
  • Hardened for production: nginx reverse proxy, Keycloak OIDC SSO, per-user allowlisting, fail-fast config checks

Result

Now live in production for one functional unit at Greenko (Company Secretarial / IMS documentation), architected to extend department-wide — with consistent document-control metadata and zero outsourcing.

FastAPI vLLM (Qwen) AWS Bedrock (Claude) React + TypeScript Docker + nginx Keycloak / OIDC Word DrawingML

Selected work

Projects

A curated set — from a production optimization POC to an M.Tech thesis on edge AI. Full detail available on request.

Greenko Group · Operations Research · Jan – May 2026

Warm-Start Optimization POC (GEMOS)

Greenko's Day-Ahead Market solver rebuilt its optimization from scratch every day, even though the problem barely changes. I designed and benchmarked two warm-start strategies — reusing yesterday's solution (D-1) and last week's same-weekday solution (D-7) — feeding prior solver states in as an MIP start.

71.3% best-case runtime cut 75% win rate (D-1) ~40–50% avg. speed-up

The harder part was the failure mode: warm starts silently broke when a contract changed the problem's structure, even though every variable bound still checked out. I traced it to a constraint the old solution never accounted for, then built the safeguard — validate first, fall back to cold start on any doubt.

Wk 1–3
Knowledge transfer, LP/MIP foundations
Wk 4–6
Advanced solver tuning, warm-start proposal
Wk 7–9
POC build, MinIO persistence, benchmarking
Wk 10–12
Automated testing, failure root-cause analysis
GurobiMIPPythonRayMinIO

M.Tech Thesis · IIT Hyderabad

KathaMitra — offline AI storytelling for kids

An Android app that turns screen time into Socratic reflection for 9–12-year-olds — narrated stories drawn from Indian epics, followed by open-ended questions the child talks through out loud. Every bit of it runs on-device.

<450ms end-to-end latency 100% local inference 4.80/5.0 pedagogical score

The tricky part was handling how kids actually talk — code-switched Hindi/English speech. I designed a "Symmetric Context" retrieval step that pulls clean English narrative concepts from Hinglish speech before querying the local vector index, then fuses that back with the child's original words so responses stay in their natural voice, grounded in verified stories rather than improvised.

Gemma 2B (quantized)MediaPipe / LiteRTsqlite-vecAndroidOn-device STT/TTS

Deep Learning · AI5100 · IIT Hyderabad · Aug – Nov 2024

Scientific Diagram Captioning

A multimodal system that reads scientific diagrams and writes their captions — combining CLIP and Vision Transformers for visual understanding with BART for generation.

73% Recall@10 4M-image SciCap subset
CLIPVision TransformersBARTPyTorch

NLP · CS6803 · IIT Hyderabad · Aug – Nov 2024

Toxic Content Detection

Binary and multi-class classifiers to flag harmful social media content for automated safety screening, fine-tuning HateBERT on the DHate dataset.

Competitive macro-F1
HateBERTTransformersNLP

Operations Management · EM6040 · IIT Hyderabad · Jan – Mar 2025

Biogas Supply Chain Optimization

A multi-period optimization model for sourcing and blending four waste feedstocks across four Telangana biogas plants — maximizing profitability across a full year of seasonal variation.

20:1–35:1 C/N ratio maintained
OptimizationOperations Research

Strategic Management · EM6010 · IIT Hyderabad · Dec 2024 – Mar 2025

Sarvam AI: Building India's Sovereign Generative AI

A strategic study of Sarvam AI's voice-first multilingual strategy — business model, competitive position, and route from seed stage to enterprise deployment, including the trade-offs around Indic-language models and data sovereignty.

Market analysisGenerative AI strategyIndia data sovereignty

Global collaborations

Applied research and strategy work alongside international student founders and researchers.

Nagoya University · Feb 2025

Tongali Overseas Project — refined startup business models and marketing strategy with Japanese student founders.

J-StarX × IIT Hyderabad · Feb 2025

Assessed India's parental-control-app market for a Japanese child-safety startup; built the reverse market-entry pitch.

Kyoto University · Sep 2025

Deep-Tech Entrepreneur Spirits — mapped urban-sustainability challenges under SDG 11 with Kyoto University students.

Shimane University · Sep 2025

Presented ongoing thesis research, exchanging perspectives on innovation and applied research.

Toolbox

Skills

Languages & Data

  • Python
  • C++
  • MySQL / PostgreSQL
  • MongoDB

AI / ML

  • LangChain / LangGraph
  • Transformers, BERT / HateBERT
  • CLIP, Vision Transformers, BART
  • LSTM, contrastive & generative modeling

Backend & Platforms

  • FastAPI, Docker
  • React, Keycloak
  • Ray, vLLM, AWS Bedrock
  • sqlite-vec, MediaPipe

Systems

  • Systems design
  • Distributed computing
  • Algorithms
  • Mixed-integer programming / OR

Always shipping, always learning

Continuous learning

Ultimate RAG — Krish Naik Academy

4 months · 16 modules · 5 projects

Alongside shipping production RAG systems at Greenko, Sumanth completed a 4-month intensive covering the full retrieval-augmented generation stack — from foundational similarity search to production deployment on Kubernetes. It closes the loop between what he builds on the job and what's next in the field.

Core retrieval

RAG foundations, LangChain, LlamaIndex, Haystack, document parsing

Advanced retrieval

Hybrid search & reranking, Self-RAG, Graph RAG, query expansion (HyDE)

Agentic & multimodal

LangGraph, ReAct agents, multimodal RAG (CLIP + Whisper), Model Context Protocol

Shipping it

Docker, Kubernetes (AWS EKS), CI/CD, evaluation & monitoring

5 end-to-end projects built, including an MCP-integrated code-review assistant and a cloud-deployed multimodal RAG system on AWS EKS.

Timeline

Experience & education

JAN 2026 — JUN 2026 · HYDERABAD

Intern, Gen AI / Operations Research

Greenko Group

Delivered two production-oriented projects in under five months — a warm-start solver POC for energy-market optimization and a production-deployed AI documentation agent — while completing an M.Tech thesis in parallel.

  • Warm-Start Optimization POC (Jan – May): up to 71.3% solver runtime reduction across 8 benchmark cases
  • GIMS AI Agent (May – Jun): production-deployed 5-pipeline document automation platform with SSO

2026

M.Tech, Techno-Entrepreneurship

IIT Hyderabad · CGPA 8.73

Coursework spanning NLP, computer vision, deep learning, strategic management, operations management, and design thinking — bridging deep technical work with product and market strategy.

FEB 2025 — SEP 2025

International research & strategy collaborations

Nagoya University · Kyoto University · Shimane University · J-StarX

Worked with student founders and researchers across Japan on market-entry strategy, urban sustainability (SDG 11), and applied AI research exchange.

2024

B.Tech, Computer Science and Engineering

Malla Reddy College of Engineering · CGPA 7.50

NOV 2024 — JUN 2025

Leadership & recognition

SPIC MACAY, IIT Hyderabad Chapter

  • Hospitality Coordinator, SPIC MACAY IIT Hyderabad Chapter and its 10th International Convention
  • Third place, International Yoga Day Competition, IIT Hyderabad

Get in touch

Let's build something that has to work in production.

Open to backend and AI engineering roles — especially where production RAG, agentic systems, or optimization meet real infrastructure constraints.