00Introduction

Dheeraj
Thuvara

AI / ML Engineer Full Stack Developer CTO & Founder

Building intelligent systems at the intersection of deep learning and scalable engineering. From CNN-BiLSTM models to production-grade Rails platforms — I ship end-to-end.

GitHub LinkedIn Resume
12+
Years Experience
5
Companies
Dheeraj Thuvara
Available for work
Scroll
01Selected Work

Projects & Experiments

Featured
01
TWO-AGENT AI AI SYSTEMS

Mini-Glia

Researcher + Supervisor · discrete-event simulator · a reproduction of the Glia paper (arXiv:2510.27176)

R S

A faithful miniature reproduction of the Glia architecture: a Researcher agent proposes, codes, and tests scheduling strategies while a Supervisor agent (with no code access) probes its reasoning and nudges it to compose ideas. Working from a discrete-event LLM-inference simulator, the agent rediscovers the paper's core insight on its own — reserve KV-cache headroom to prevent the evictions that thrash a packed GPU cluster. Watch a captured run, or run the two agents live and unscripted. An honest toy: real mechanism, illustrative numbers, no GPU.

2
Collaborating agents
~85%
Latency cut vs baseline
live
Unscripted agent runs
0
GPUs required
Hypothesize → Sandbox → Simulate → Analyze → Supervise
Python FastAPI React LiteLLM Gemini SSE streaming Docker GitHub Actions
02
MULTI-AGENT AI CLINICAL RESEARCH

Clinical Research Agent

LangGraph · Gemini · Qdrant · MCP · End-to-end evidence synthesis

A multi-agent system that answers clinical questions by orchestrating six specialized agents — PICO decomposition, semantic cache lookup, multi-source search, relevance screening, evidence synthesis, and deterministic citation fact-checking. Queries PubMed, ClinicalTrials.gov, and the FDA in parallel, then produces a GRADE-Lite quality-rated report with verified citations. Includes a custom MCP server exposing clinical tools to Claude Desktop and other MCP clients, plus full Langfuse observability across every agent run.

6
Agents in pipeline
~20s
End-to-end latency
3
MCP tools exposed
$0.002
Cost per query
PICO → Cache → Search → Screen → Synthesize → Fact-check
LangGraph FastAPI React Gemini 2.5 Qdrant MCP Langfuse Docker GitHub Actions
03
END-TO-END MLOPS HEALTHCARE AI

ECG Arrhythmia Detection

CNN-BiLSTM · MIT-BIH Arrhythmia Database · Production-grade pipeline

A production-grade MLOps pipeline detecting cardiac arrhythmias from ECG signals. Covers the full ML lifecycle — data versioning with DVC, experiment tracking with MLflow, containerized inference with FastAPI and Docker, live Gradio demo, GitHub Actions CI/CD, and Evidently AI drift monitoring. Achieves 0.98 macro F1 and 99% weighted accuracy across 5 arrhythmia classes.

0.98
Macro F1
0.91
F1 on APC (hardest)
5
Arrhythmia classes
110K
ECG beats
MIT-BIH Data → DVC → CNN-BiLSTM → MLflow → Docker → FastAPI → Evidently
PyTorch CNN-BiLSTM FastAPI MLflow Docker Gradio DVC Evidently AI GitHub Actions
02AI Applications
Machine Learning Projects 3 projects
02 — 01
Traffic Signal Analysis

YOLOv8-powered traffic density analysis that optimizes signal durations for smarter urban traffic management.

02 — 02
Sentiment Analysis

Decode the emotional tone of any text instantly using a spaCy NLP model — turns paragraphs into emotions.

02 — 03
Text Summarizer

Instantly summarize any text into one paragraph, two paragraphs, or bullet points using an LLM.

03Full Stack
Ruby on Rails Projects 2 projects
03 — 01
WishBoard

Social platform on Ruby on Rails to share, track, and monetize wishes with instant messaging.

03 — 02
Cocociti — Rental Aggregator

Full-featured rental aggregator with listing, search, and payments across 30 categories and 10,000+ products.

04Career History

Experience

Download Resume
Software Engineer & ML Platform — Apnimed Inc.
09/2025 – 03/2026
— Built scalable Python inference service exposing model predictions via REST API.
— Developed Caliper — custom internal tool for visualizing sleep stage transitions.
— Architected MLflow experiment tracking pipeline across 300+ GPU training hours.
— Trained hybrid CNN-BiLSTM for multi-class sleep stage classification on RTX 4070 and A100.
Technical Manager — Encora Innovation Labs
03/2019 – 12/2024
— Led development for WebAdMIT managing a team of 7 using Agile and Scrum.
— Improved accessibility from 20% to 95% with 508 standard and NVDA support.
— Managed 100,000+ college applications per minute with full system stability.
— Optimized PostgreSQL across 20M+ applicant records powering large-scale dashboards.
Senior Software Developer — GrabOnRent
11/2017 – 03/2019
— Built Rails modules for rental platform with 300 products and 5,000 SKUs.
— Migrated frontend to ReactJS integrated with Rails REST APIs.
— Implemented Elasticsearch improving search performance by 40%.
Founder & CTO — Cocociti Int. Pvt. Ltd.
10/2015 – 11/2017
— Built social media platform with instant messaging for wish sharing and monetization.
— Led team of 5 overseeing coding, server ops, DB management, and UI design.
— Cut bounce rate 20% and increased engagement 25% using analytics tools.
Systems Engineer — Tata Consultancy Services
02/2013 – 10/2015
— Developed modules for fitness tracking and CSR platforms.
— Migrated databases from MySQL to PostgreSQL for enhanced performance.
05Get in touch

Let's build
something great

Open to full-time roles and freelance in ML Engineering and Full Stack Development.

Email me LinkedIn GitHub Resume
[email protected]