Most engineers ship code. I ship research that ships back. Author of QEMA-G, a quantum-enhanced memory framework for graph AI published on Zenodo, and of The Coin in the Air, a book on quantum machine learning out now on Amazon. VP of R&D at AI Skunkworks, the graduate research org I rebuilt at Northeastern. Four years of production engineering, now building the layer where multi-agent systems, RAG, and reinforcement learning converge into LLMs that do useful work in the real world.

Multi-agent systems with statistically validated evaluation, RL-driven orchestration, fine-tuned models, and production data platforms. Each one shipped, tested, or published - not a tutorial follow-along.
Citation-grounded multi-agent review: three specialised agents ground every finding in a 70-passage OWASP Top 10 2025 / CWE knowledge base, with an adversarial evaluator over each claim. 97% fewer hallucinated findings (FPR 0.789 → 0.111, McNemar's p = 0.0312) vs. a single-prompt baseline - same model, same prompts, different architecture. Ships as CLI, REST API, SARIF 2.1.0, Docker/Helm.
Layers a UCB-1 contextual bandit (per-code-type agent configuration) and REINFORCE policy gradient (2,853-param MLP for finding synthesis) onto the CodeSentinel pipeline, so review quality improves with developer feedback: +28.1% mean reward vs. random baseline across 200 episodes × 5 seeds (95% CI [25.6%, 30.7%]).
3-agent CrewAI + n8n hybrid with a 940-line custom tool querying 16 academic databases in parallel. Achieved 85% source relevance (+27 pts over web-only baseline), 90% factual consistency, and 95% citation completeness across 45 evaluation runs - peer-validated at Cohen's κ = 0.74.
An all-seeing intelligence layer: pluggable ingestion channels normalized to one schema, persistent memory, a temporal knowledge graph, watchlists with alerts, and cited brief synthesis. Exposed as CLI, REST API, and MCP server so both humans and agents can consume it - with a dependency-free offline core.
Theoretical framework integrating quantum memory architectures with graph neural networks to attack the memory-wall problem in large-scale GNN training. Author: Aravind Balaji · Co-author: Prof. Nik Bear Brown. Published on Zenodo, with a Qiskit reproduction suite covering noiseless and IBM-Brisbane noise models.
Tracks the live public space catalog (~16,000 objects): parallel SGP4 propagation in Rust (full catalog in milliseconds), three-stage conjunction screening with collision-probability estimation, an axum REST API, and a real-time 3-D operations console. Screens 16k objects × 6h in about a minute on a laptop.
LLMs win IMO gold yet still ship confidently wrong math. Verimathix is the trust layer: step-level verification with first-error localization (SymPy symbolic + numeric probing), natural-language → Lean 4 autoformalization, reference numeric kernels, and a step-grading benchmark harness with CI.
Fine-tuned TinyLlama-1.1B with QLoRA on the Spider dataset for NL-to-SQL generation - 100% improvement on SELECT-clause generation against the base model. Shipped with a ChromaDB RAG retrieval layer over schema metadata and a Gradio demo.
Offline-first clinical knowledge-graph and decision-support platform with FHIR / HL7 / C-CDA connectors for US + India. Graph-aware RAG: Neo4j retrieval, structured reasoning, and natural-language explanations of clinical relationships over Snowflake-hosted data.
Fault-tolerant ingestion pipeline, deployed live on GCP: Cloud Run API with instant 202 acks → Pub/Sub buffering → workers with retries and dead-lettering, per-tenant Firestore isolation, PII redaction, and load testing.
30+ open-source repositories across agentic AI, domain intelligence platforms, and production systems.
Intelligence-fusion platform with a citation-grounded AI analyst copilot - ADS-B/AIS/GDELT fusion, entity resolution, anomaly detection.
Java · HealthcareClearAuthCMS-0057-F prior-authorization platform - FHIR-native PAS intake, auditable adjudication, SLA enforcement, CMS PA metrics.
Python · ClimateCarbonoscopeOpen climate intelligence: auditable GHG accounting, Climate TRACE analytics, grid carbon forecasting, disclosure parsing.
Python · AgTechAgroCortexAI cortex for agriculture: LLM benchmark, plant-disease vision, satellite field analytics, citation-grounded AI agronomist.
Python · Civic TechCiviSynthPolitical intelligence platform: narrative tracking, RAG fact-checking with evals, legislative intelligence, election forecasting.
Python · RAGOmniCanonCitation-verified AI for sacred texts - multi-faith corpus, zero-hallucination RAG, TheoBench LLM benchmark, canon-aware OCR.
Python · Sci-MLPhysWeavePhysics simulation meets ML: differentiable simulation, neural PDE surrogates, stochastic finance, quantum circuits, molecular dynamics.
Python · ChemistryAlcheMindDe novo molecular discovery: RDKit + genetic generation + ML property prediction + FastAPI, end to end.
Python · RoboticsEveryCamEvery camera is a robot teacher - privacy-first, robot-ready embodied data from everyday cameras.
Go · COBOL · FinTechQuantumTellerThe open-source ATM stack: ISO 8583 switch in Go + COBOL core banking + roadmap to ML fraud scoring and post-quantum crypto.
TypeScript · Full-StackServiQue v2Production-grade service booking platform - Node + TypeScript + Express + Prisma API, React SPA, JWT/RBAC, tests, Docker, CI.
Python · Edge AISmart TrolleyEdge-AI shopping trolley on Raspberry Pi - VGG16 → MobileNetV3 TFLite INT8 produce recognition, load-cell weighing, Firebase billing.
React · 1st PlaceDNATE MSL Practice Gym1st place, MGEN Hackathon 2025 - AI conversation simulator for Medical Science Liaisons (React, Node.js, OpenAI API), built end-to-end in 48 hours.
Jupyter · RLLLM Agents & Deep Q-LearningLLM agents and DQN experiments on Atari games - classic RL meets modern agent tooling.
MERN · HealthcareCareHubHospital management system - MERN stack with JWT authentication and role-based access control.
HTML · GamesCheese Chase ArcadeTen arcade games + a 10-chapter story mode in a single HTML file. Zero dependencies.
HTML · AuthorIdeas You Can PictureOfficial website for my book series explaining quantum computing and AI with zero equations.
GitHub…and more →The full list - 30+ repositories, updated constantly.
Four years across healthcare technology, cloud migrations, and SaaS reliability before grad school - shipping in production, not just classrooms.
Led a 4-person team to first place, building an AI conversation simulator for Medical Science Liaisons (React, Node.js, OpenAI API) end-to-end in 48 hours.
Deep on AI/ML and Python; fluent across the data, cloud, and full-stack layers needed to ship end-to-end systems.
Northeastern University, Boston · GPA 3.90/4.0 · Expected Dec 2026
BITS Pilani, India
SRM Institute of Science & Technology, Chennai
Aravind Balaji, Nik Bear Brown · Northeastern University. A theoretical framework with feasibility analysis, published on Zenodo - plus an open Qiskit simulation suite reproducing every numerical result under noiseless and IBM-Brisbane noise models.
Contributor to Prof. Nik Bear Brown's textbook on agentic systems: retrieval-augmented generation as an evidence layer for LLM agents, evaluation methodology for citation faithfulness, and production case studies.
Book 1 of Ideas You Can Picture - a six-book series explaining quantum computing and AI with clear pictures and not a single equation.
Quantum computing may be the most exciting idea of our time - and the most badly explained. Through one simple, unforgettable image - a coin spinning in the air - you'll finally understand superposition, entanglement, interference, and the new kind of AI they're being turned toward. 27 commute-sized chapters. Zero math. No background required.
Five more books in the series are on the way, covering quantum cryptography, quantum hardware, and modern AI.
Narrative nonfiction on AI engineering, quantum computing, and the human side of the systems we build - published on Substack and Medium.
The mathematics of signal processing is the mathematics of attention mechanisms.
Read on Substack →
Quantum ComputingThe Quantum Librarian: QEMA-G ExplainedA plain-English walkthrough of the QEMA-G framework - told as a story rather than a paper.
Read on Substack →
Hardware & AIInside AWS TrainiumWhat custom AI silicon actually buys you, and why the chip wars have only just started.
Read on Substack →
Beyond the code: I also produce independent music with AI tools - same loop as engineering: idea → prototype → critique → ship.
Most candidates pick a lane - production engineering or research, models or systems, theory or shipping. I'm fluent in all three, which is the rarer profile, and the one that makes AI initiatives actually move.
Open to: AI / ML Engineering · Applied Research · Multi-Agent Systems · Data & ML Platforms · Full-Stack with an AI layer
When: Open to Co-op / Internship & Full-time roles - starting immediately. No visa sponsorship required.
How: In Person, Remote or Hybrid. Where: Boston, Anywhere in the United States and around the world.