Lohith Devaramane

SWE @ Rakuten Symphony · FDE

Hey, I'm Lohith.

I help teams ship AI into production the way I've shipped telecom software for four years — forward deployed, production first. Automation engineer at Rakuten Symphony, now completing an 8-month FDE intensive.

View deploymentsAsk my copilot →
Lohith Devaramane Shivashankara

Live deployment map

initializing deployment network…

Every node is a real engagement or a slotted build from my 8-month FDE intensive. The network grows as I ship.

06

deployed systems

02

in progress

23

queued in pipeline

4+

years in production

Flagship engagements

Written like engagement reports, not project blurbs

ENG-030deployed

Split — RSVP links and fair cost splits for group games

Paste a court booking, post one link in the WhatsApp group, friends tap “I’m in”, and the cost splits itself — each person gets their share with a one-tap UPI link, balances carry across games, and reminders go out until it’s settled.

Next.jsSupabasePostgres RLSGoogle Apps ScriptVercel
Open case study →
ENG-029deployed

Bellcourt Review Copilot — agentic RAG for prior authorization

A review copilot for a health-plan administrator: picks the policy version in force, checks every criterion against the record with quoted evidence, and leaves the decision to a person — 117 of 120 audited cases right, zero denials possible by design.

PythonAgentic RAGGemini 2.5 FlashVector indexVercel FunctionsSupabase
Open case study →
ENG-027deployed

SetuHaul — dock scheduling for late freight

A driver reports a delay in Hinglish and every party — warehouse, carrier, customer — has the revised plan before the truck arrives. Double-booking is impossible by database constraint.

PythonFastAPILangChainGemini 2.5 FlashPostgreSQLOR-Tools CP-SAT
Open case study →
ENG-028deployed

Lead Intelligence System — rubric-first lead qualification

1,200 inbound leads a month, 60 worked. A CLI that scores every lead against an explicit rubric, sends batches of twenty to Gemini for reasoning and a personalised first message, and hands sales a ranked queue — with the close calls routed to a human, not guessed.

PythonLangChainGemini 2.5 FlashYAML rubricBatch LLMRetry + fallback
Open case study →
ENG-000deployed

RAN performance test automation

Production KPI-testing automation for live radio access network setups at Rakuten Symphony — 70% less manual effort, 50% faster suites.

PythonSeleniumJenkinsMultithreadingLinux
Open case study →
ENG-001deployed

Slicematic — AI operations copilot for a pizza outlet

Order management, live tracking, and owner-facing AI insights that turn daily order data into margin-growth decisions.

Next.jsAI insightsVercel
Open case study →
ENG-002in-progress

This site's copilot

A retrieval-grounded assistant trained on my deployments and case studies — with a designed failure mode when the model is unreachable.

Next.jsRetrievalGemini APIVercel Functions
Open case study →

Track 01 — Technology

Build the system

Industrial RAG, agents and multi-agent orchestration, MCP servers, fine-tuned SLMs, full-stack AI apps, cloud deployment — on top of four years of Python automation, CI/CD, and multithreaded systems in telecom production.

Track 02 — Consulting

Land the system

Discovery interviews, MECE problem decomposition, business-to-technical translation, PoV→PoC→MVP scoping, and stakeholder management — the half of forward deployment that never shows up in a GitHub repo. See field analysis →

On-site assistant

Don't read the whole site. Interrogate it.

The copilot answers questions about my experience from this site's content, with citations — and it's a documented engagement itself (ENG-002).

Open copilot