Mahesh Balan
mahesh@pravici.com · mahesh.balan@cgu.edu
Enterprise delivery leader, founder, and doctoral researcher. I focus on systems that ship—GenAI and classical platforms, Enterprise Blockchain Technology, health wallets, federated learning, and teaching—so ideas become measurable outcomes for organizations and patients.
Executive summary
Enterprise AI transformation leader and former Oracle/Siebel Technical Director with 30+ years building and scaling revenue-generating consulting organizations and delivering enterprise systems across Fortune 500 and public-sector clients. I am a leader in applying blockchain technology to enterprise software: Pravici’s coalition loyalty platform is built on the invention claimed in my U.S. patent US11854038B1 — a multi-party, multi-point-type decentralized loyalty system using a permission-based distributed ledger for promotion, point issuance, and redemption. Founder & CEO of pravici.com, scaling consulting to $2M–$4.5M ARR and launching AI-enabled SaaS platforms. Hands-on with GenAI and LLM deployments (ChatGPT API, RAG, vector databases, evaluation, monitoring, governance) and experienced leading customer-facing engineers through workshops, structured adoption, and post-go-live optimization. Published IEEE researcher and U.S. patent holder; doctoral candidate with intensive AI research, model deployment, and teaching at Claremont Graduate University. I am a Student Fellow of the AI for Humanity Lab at CGU, focused on responsible, human-centered AI.
Core themes: AI success & adoption · enterprise delivery · customer enablement & workshops · GenAI / LLM & RAG · evaluation & observability · responsible AI & data privacy · APIs · PostgreSQL / pgVector · cross-functional leadership (sales, product, engineering).
Doctoral research: health wallets & federated learning
Doctor of Technology (AI & Healthcare), Claremont Graduate University — expected graduation Spring 2027. Two-plus years of intensive AI research, model deployment, and responsible AI study. I am a Student Fellow at the AI for Humanity Lab, where we advance AI grounded in human values, ethics, and real-world impact.
Key AI coursework completed: AI for Digital Transformation, AI for Digital Transformation Practicum, Natural Language Processing, Generative AI and Applications, Deep Learning & Computer Vision, Machine Learning for Healthcare, Persuasive Technology & Ethics, Introduction to Use-Inspired Research, Advances in AI and Quantum Computing.
MyWellWallet — a health wallet I am developing as part of my doctoral work. It now runs the latest MedGemma 4B-parameter model locally for natural-language interaction, custom-tuned for medical terminology, alongside a Model Context Protocol (MCP) client connected to a FHIR MCP server I am building (mcp-fhir-server.com). The goal is a patient-centered, 360-degree view of health data with conversational access throughout.
My research extends MyWellWallet to participate in decentralized model building through federated learning: enabling collaborative improvement of privacy-sensitive models across participants without pooling raw PHI, while aligning incentives and trust with on-chain accountability where appropriate.
Federated Learning with SpEG Scoring
Title: FLAI Protocol: Decentralized Federated Learning with On-Chain Rewards and sPEG-Based Contribution Scoring
Description: A blockchain-coordinated federated-learning protocol pairing decentralized training with tamper-evident ledger records, on-chain reward mechanics, and sPEG-style contribution scoring so participant value is measurable and equitable.
Links: ICCE conference paper PDF · EDAS session listing
FedLoRA Fine-Tuning with TurboAggregate Secret Sharing
Title: Decentralized Federated LoRA Fine-Tuning for On-Device LLMs: A Blockchain-Based Approach
Description: FedLoRA-based fine-tuning for on-device large language models, coordinated via a blockchain layer and secured with TurboAggregate aggregation and secret-sharing–style cryptography to shrink communication overhead while resisting inference from individual updates.
Link: BCCA accepted paper PDF
Device-Tier Heterogeneity in Federated Health-Risk Scoring
Title: Device-Tier Heterogeneity in Federated Health-Risk Scoring: A Measured Study on Consumer Edge Nodes
Description: Measures how device-tier feature heterogeneity—different home medical sensors across households—affects federated diabetes-risk scoring on consumer edge nodes. Using real U.S. NHANES data on CPU-only hardware, shows that training only on features every home shares drops 15.6 AUROC points, while a tier-aware federated averaging rule recovers 13.9 points, converges in 2 rounds instead of 26, and yields per-tier marginal contributions suitable for on-chain incentives—with wall-clock, energy, and uplink costs measured on the node itself.
Link: AIBThings accepted paper PDF
Clinical LLM Deployment on Consumer CPUs
Title: Capacity over TFLOPS: Memory-Centric Deployment of Large Clinical Language Models on Consumer CPUs
Description: Treats the edge node as capacity-enabled (system RAM) rather than GPU-limited for large clinical language models—measuring INT4 27B-model footprint, CPU-based inference latency, energy, and federated LoRA communication cost on commodity hardware with synthetic FHIR prompts, showing a privacy-preserving path for on-premises clinical AI.
Link: ICCE Berlin accepted paper PDF
Lightweight Biometric Drift Detection on Edge Nodes
Title: Lightweight, Label-Free On-Device Biometric Drift Detection on a Consumer Health Edge Node
Description: Treats biometric drift detection—flagging sustained regime shifts in multi-modal wearable streams—as a lightweight, label-free task on consumer-owned edge nodes. Evaluates detectors from a closed-form cross-channel Mahalanobis test to lightweight convolutional models on real Fitbit and Empatica datasets; sustained mean-level shifts are reliably detected (AUROC 0.92–1.00), with simple classical methods Pareto-optimal on a Galaxy S10 (0.58 ms inference, ~0.58 mJ per check) and continuous monitoring costing on the order of 10−4% of battery per day.
Link: HMSS accepted paper PDF · HMSS 2026 conference
Consumer-Owned Edge Health Node
Title: A Consumer-Owned Edge Node for Verifiable, Privacy-Preserving Personalized Health
Description: A consumer-owned home edge appliance that aggregates wearables, home medical devices, and EHR data into local FHIR, runs on-device health-risk analytics, and crosses the privacy boundary only with cryptographically verifiable health indicators and compliance proofs (zk-SNARKs)—so insurers, clinicians, or researchers can trust outcomes without receiving raw PHI.
Label-Correlated Censoring in Federated Waveform Learning
Title: Filter Gently, Then Drop Bits: Label-Correlated Censoring in Federated Waveform Learning
Description: Shows that on-device signal-quality filtering—standard practice before uploading wearable waveforms—introduces selection bias in federated learning because quality is correlated with physiology (e.g., motion degrades PPG while raising heart rate). Discarding windows by quality costs far more accuracy than discarding the same volume at random; under tight uplink budgets, lossy transmission of every window outperforms selective transmission of only “clean” ones.
Signals for Bitcoin Market Intelligence
Theme paper: Reddit Versus News: Signals for Bitcoin Market Intelligence
Full title: Institutional vs. Retail Information Channels for Cryptocurrency Market Intelligence: A Comparative Analysis of Signal Quality and Decision Support Implications
Description: Work from the OasisCoin research line comparing alternative information channels—from social (Reddit) versus news-derived signals through institutional disclosures versus retail-oriented feeds—for cryptocurrency market intelligence, with implications for trader decision-support quality.
Links: AMCIS accepted paper PDF · AMCIS 2026 program
Broader interests include agentic technology and user ownership of data and consent, verifiable credentials, decentralized web nodes, self-sovereign identity, and cryptography (threshold signatures, differential privacy, homomorphic encryption).
Entrepreneurship & products
- Founder & CEO · pravici.com · 2013–present Scaled an enterprise consulting portfolio to $2M–$4.5M ARR; owned P&L, pipeline, executive relationships, and delivery. Clients have included General Motors, Vauxhall Motors, Opel, Groupe PSA, Zebra Technologies, Quest Software, NXP Semiconductor (Freescale), and realtor.com. Built and deployed Pravici SaaS: Pravici TLP, RnR, WPX, and PocketCred — verifiable credentials for health, lifelong learning, and more using SMART on FHIR.
- Fractional CTO · MyNuGen.ai · 2025–present Leading ChatGPT-based retrieval-augmented generation (RAG) architecture for the platform’s agentic AI college admissions and coaching stack (OpenAI API, pgVector, similarity modeling, evaluation workflows).
Additional product engineering (selected)
- Bespoke blockchain loyalty (Ethereum, Hyperledger Fabric) for an automotive OEM.
- Verifiable credentials and NFT-based learning and employment records using OpenBadges 3.0 and ERC-721 / ERC-1155, with OpenZeppelin.
- NFT marketplace for carbon credits on Polygon using Infura and Pinata.
Teaching & speaking
Claremont Graduate University
- Teaching Assistant: AI for Digital Transformation (IST 601)
- Teaching Assistant: AI Practicum (IST 697)
- Teaching Assistant: Machine Learning for Healthcare (IST 347)
- Teaching Assistant: AI-Powered Enterprise Systems and Intelligent Supply Chains (IST 314)
- Teaching Assistant: AI for Humanity (TNDY 409D)
- IEEE ICCE 2026 — Conference tutorial
Harvey Mudd College / AXL Conference 2026
- May 21, 2026 — Lightning talks on the societal impacts of GenAI, (Re)Imagining Liberal Arts & STEM Education in the Age of GenAI (AXL Conference 2026), Harvey Mudd College, Claremont, CA. Talk: Local LLMs, Global Impact: Reimagining AI Ethics Through Patient-Owned Health Wallets — MyWellWallet as a case study for local/medical LLMs, patient-mediated FHIR, federated learning, and the environmental & ethical tradeoffs of on-device vs. cloud GenAI. Session page · Slides (Google)
Industry & community (verifiable credentials, privacy, health)
Earlier experience
- Technical Director · Oracle / Siebel · 2000–2013 Led 30+ enterprise CRM deployments; partnered with sales and product; coached cross-functional engineering teams. Sample clients: Toyota Financial Services, Nissan Motor Acceptance, Hyundai Motors, Toshiba; Gilead Sciences; Motorola SPS; Cymer (ASML); Honeywell; McAfee; Nationwide Financial Services; Peregrine Systems; Jenny Craig.
- SAP BASIS Manager & Technical Systems Specialist · City of Phoenix · 1990–2000 SAP ERP for City of Phoenix Finance: BASIS operations, reliability, security, disaster recovery, and compliance. Phoenix Police: mission-critical 911 dispatch, criminal history, and AFIS systems — high availability, fault tolerance, data integrity, and 24×7 operations in regulated environments.
Education
- Doctor of Technology candidate (AI & Healthcare), Claremont Graduate University — expected Spring 2027
- Master of Science, Computer Science, Arizona State University, 1996 — GPA 4.0/4.0
- Bachelor of Engineering, Mechanical Engineering, Birla Institute of Technology & Science, 1986 — GPA 9.6/10, first in class
Patent, papers & articles
- U.S. Patent US11854038B1 — Multi-party decentralized loyalty on a permissioned ledger
- AMCIS 2026 (Reno NV, accepted) — Institutional vs. Retail Information Channels for Cryptocurrency Market Intelligence (PDF) · OasisCoin (GitHub) · Conference program
- IEEE ICCE 2026 (Dubai) — FLAI Protocol (full PDF)
- IEEE ICCE 2026 (Berlin, accepted) — Capacity over TFLOPS: Memory-Centric Deployment of Large Clinical Language Models on Consumer CPUs (PDF)
- IEEE BCCA 2026 (Barcelona, accepted) — Decentralized Federated LoRA Fine-Tuning for On-Device LLMs: A Blockchain-Based Approach (PDF)
- IEEE AIBThings 2026 (Central Michigan University, Sep 5–6, accepted) — Device-Tier Heterogeneity in Federated Health-Risk Scoring: A Measured Study on Consumer Edge Nodes (PDF)
- HMSS 2026 (Barcelona, Nov 17–20, accepted) — Lightweight, Label-Free On-Device Biometric Drift Detection on a Consumer Health Edge Node (PDF) · Conference site
- IEEE Consumer Electronics Magazine (submitted) — A Consumer-Owned Edge Node for Verifiable, Privacy-Preserving Personalized Health: consumer-owned home edge node, local FHIR, on-device analytics, and zk-SNARK verifiable health indicators without disclosing raw data
- IEEE SCSETech 2026 (Pomona, CA, Dec 1–3, submitted) — Filter Gently, Then Drop Bits: Label-Correlated Censoring in Federated Waveform Learning: on-device signal-quality gates can bias federated waveform models when quality correlates with the label; under tight uplink budgets, lossy transmission of every window can beat selective “clean-only” uploads
- Advancing employee behavior analysis through synthetic data (ABMs, GANs, statistical models)
- Unveiling employee behavior dynamics through Wolfram cellular automata
- If data is the new oil, where is your health data?
- “Member as API” — Interoperability rule & verifiable credentials
- Loyalty and health — improving decisions with a nudge