PARITOSH KULKARNI
ASSET FAULT PREDICTION · INDUSTRIAL AI STRATEGIST · IMPLEMENTATION & EVALUATION EXPERT
I help industrial manufacturers, power & energy utilities, and private equity portfolio companies move AI from stalled pilots into auditable, enterprise-wide production. Over 7+ years as a founding engineer, CTO, and principal AI architect, my work sits at the intersection of industrial asset fault prediction, real-time ERP/MES/historian data engineering, and rigorous statistical & agent evaluation.
Whether streaming 328 Oracle ERP/MES schemas into Palantir Foundry at T1 Energy ($1B solar manufacturer), taking clinical AI from a 0→1 startup acquisition to enterprise cohort stratification at Grifols ($7B+ biopharma), building competing-risk transformer fault prediction models for Tier-1 California utilities (PG&E / SCE track), automating BOM and procurement operations at Achilles, or transforming PE-backed portcos at Attian AI, I gate every deployment on baseline (“Model 0”) reproduction, out-of-time backtests, and unit-economics due diligence.
Asset Fault Prediction & Reliability
Competing-risk survival modeling, Dissolved Gas Analysis (DGA), PI historian thermal/load stress, and PoF × CoF ($/asset/year) replacement optimization across substation power transformers, grid-scale battery cells, and plant machinery.
Pilot-to-Enterprise Implementation
VPN-secured Change Data Capture (Oracle LogMiner, SAP S/4HANA, enova365, MES, SCADA) into Snowflake, Azure PostgreSQL, and Palantir Foundry—plus autonomous CAD/DXF BOM and procurement agents at Achilles.
PRODUCTION DEPLOYMENTS & CASE STUDIES
Selected engagements across electric utilities, solar & battery manufacturing, biopharma, and private equity portfolio companies.
Substation Power Transformer Fault Prediction & Rate-Case Risk Intelligence
Subcontractor AI & Data Lead (PKulkarni AI Builders × GridSure Innovations) · Tier-1 California Electric Utilities (PG&E & Southern California Edison Track)
Operational Challenge
Existing utility transformer health and risk models lived as static spreadsheet snapshots that could not ingest longitudinal Dissolved Gas Analysis (DGA) lab histories, SAP PM notifications (IW29/IW39), or PI historian load/temperature traces—and conflated proactive age-based replacements with true in-service dielectric or thermal failures.
Implementation Architecture
- Codified the utility’s existing spreadsheet risk calculation ("Model 0") into an auditable software baseline reproducing historical outputs within ±0.5% before introducing any ML features.
- Linked 10–20 years of SAP PM equipment records, DGA gas ratios & rate-of-change trajectories, oil quality tests (BDV, moisture, furans), ambient heat-storm weather, and OIS outage logs across a one-way data diode with SHA-256 manifests.
- Trained competing-risk survival models separating true in-service failures from proactive planned retirements across 1-, 3-, and 5-year horizons, paired with Consequence-of-Failure (CoF) models to rank fleets by expected annual risk ($/asset/year) and Risk-Spend Efficiency.
- Deployed an air-gapped 8× NVIDIA B200 on-prem cluster running Command A+ for grounded rate-case and inspection narratives (zero invented numbers) and Inkling-Small for nameplate OCR and infrared/test-sheet extraction.
SR 11-7-aligned model validation: strict out-of-time historical backtesting (trained on pre-2020 data, evaluated on 2020–2025 failures) requiring statistically significant improvement (95% bootstrap CI) over "Model 0" on Uno’s C-index, time-dependent PR-AUC, top-decile failure capture, and calibration curves—with human Model Owner sign-off required before any retrained release.
Building an AI-Ready Modern Data Lifecycle for a $1B Solar Manufacturer
Staff Software Engineer / Industrial Data & AI Lead (Featured on Achilles & Attian AI) · T1 Energy — U.S. Solar Technology Manufacturer ($1B Revenue, 5GW Annual Production)
Operational Challenge
Production and inventory telemetry across a 5GW solar manufacturing footprint was locked behind a FortiClient VPN across 328 undocumented Oracle 19c ERP schemas, SAP modules, and legacy MES databases—forcing analysts to spend 40+ hours/week on manual CSV exports and leaving plant planners with 24+ hour stale reports.
Implementation Architecture
- Engineered an automated VPN-tunneling ingestion and schema discovery pipeline cataloging all 328 Oracle 19c schemas and activating transaction-log Change Data Capture (Oracle LogMiner CDC) into an Azure PostgreSQL event store in under 30 minutes end-to-end.
- Architected a unified EventBus schema agent with cryptographic content hashes, automated schema-drift reconciliation, and containerized self-healing Docker / GitHub Actions pipelines achieving ≥99.5% nightly reliability.
- Connected curated operational streams directly into Palantir Foundry and executive Tableau dashboards, and deployed LLM-powered Mandarin-to-English translation for Chinese supplier technical PDFs and bills of lading.
100% row-level and cryptographic hash reconciliation between source Oracle/MES transaction logs and target analytical stores; <1% pipeline error rate; cut raw-material safety buffers by 1 full day and established the real-time telemetry foundation for equipment fault prediction and demand forecasting.
Unified ERP-MES Operations & Cell-Level Fault Traceability for Grid-Scale Storage
Industrial AI & Data Architect (Featured on Achilles & Attian AI) · Grid-Scale Battery Manufacturer ($500M Revenue, 2GWh Annual Capacity)
Operational Challenge
A 2GWh grid-scale battery manufacturer had zero automated link between SAP S/4HANA and its proprietary shop-floor MES and cell testing cyclers. Inventory counts lagged by 72+ hours, quality engineers could not trace defective cells back to raw-material lots, and EU Battery Regulation 2027 passport mandates loomed.
Implementation Architecture
- Architected a real-time event backbone using Apache Kafka and Debezium CDC to stream bi-directional state changes between SAP S/4HANA, shop-floor MES, and high-frequency cell testing databases into Snowflake.
- Built end-to-end cell-level genealogy tracking every battery cell from raw-material lot and electrode coating through module assembly, formation cycling, and pack shipment.
- Implemented automated anomaly detection across cell formation and testing telemetry to flag early capacity fade and internal-resistance faults before pack integration.
Delivered 100% cell-to-pack quality traceability across 2GWh capacity, reduced order-to-delivery cycle time by 35%, and compressed customer status visibility from 3+ days to under 5 minutes within a 16-week deployment.
Achilles: Outcome-Gated Agentic Operations, CAD/DXF BOM Automation & KSeF XML Attribution
Co-Founder & CTO (Noein Labs / getachilles.ai) · Mid-Market Industrial, Custom Lighting & Discrete Converter Manufacturers (US & Europe)
Operational Challenge
Mid-market manufacturers lose weeks to manual quoting, BOM creation from bespoke CAD/DXF architectural drawings, and 10-person invoice approval chains—while generic AI chatbots fail because plant data is split across legacy ERPs (enova365, SAP, Oracle), engineering folders, and non-English supplier PDFs.
Implementation Architecture
- Built Achilles (getachilles.ai), an outcome-gated operational AI platform spanning three core modules—Catalog, Order, and Fulfill—that connects to existing ERP/MES tables, CAD/DXF drawings, PDF specs, and supplier emails without an ERP rip-and-replace.
- European Architectural Lighting ($45M manufacturer, 10,000+ SKUs): Engineered a multi-pass CAD/DXF parsing pipeline (ezdxf deterministic geometry extraction + AWS Bedrock Claude Vision for technical PDFs) that matches custom luminaire profiles against historical product families and auto-proposes 80–90% of standard BOM lines.
- Discrete/Converter Manufacturing ($35M Polish manufacturer, 2,000+ monthly invoices): Built an automated KSeF XML invoice attribution engine integrated with enova365 ERP + WMS that auto-verifies ≥60% of invoice lines and routes only low-confidence exceptions to a single reviewer.
- Implemented Monday 06:00 AM Prioritized Exception Queues and intermittent-demand SKU forecasting (Croston/TSB, hierarchical reconciliation, and conformal prediction intervals) for long-tail industrial components.
Every enterprise deployment is governed by contractual SOW acceptance gates: 80–90% BOM line auto-proposal accuracy on held-out engineering orders, ≥60% touchless invoice line verification with calibrated confidence thresholds, and 12–18 month rolling-origin forecast backtests.
From 0→1 Startup IP Acquisition to Enterprise Cohort Stratification at Grifols
Principal AI Engineer (Grifols / Alkahest) & Founding Engineer (Teal Omics — IP Acquired by Grifols) · Grifols — Global Biopharma & Plasma-Therapeutics Manufacturer ($7B+ Revenue)
Operational Challenge
Translating early-stage multi-omics discovery pipelines (UK Biobank, PPMI, Stanford Wyss-Coray lab research) and millions of noisy longitudinal EHR records into a governed, multi-indication enterprise platform inside a global biopharma manufacturer—where manual clinical cohort labeling took months and cost hundreds of thousands of dollars per indication.
Implementation Architecture
- At Teal Omics (Founding Engineer): Built the 0→1 cloud HPC and multi-omics discovery platform from a blank repo across 30+ biomedical databases (GWAS, Mendelian Randomization, pQTL, and an LLM+RAG "AI-Scientist"), enabling a $1.1M+ raise and the acquisition of Teal Omics’ plasma-proteomics IP by Grifols.
- At Grifols / Alkahest (Principal AI Engineer): Architected and deployed CHRONOS-PD, an enterprise "AI Neurologist" running on Snowflake and AWS Bedrock (Anthropic) over millions of longitudinal patient EHR records to stratify Parkinson’s disease progression.
- Designed a modular, indication-agnostic annotation and evaluation harness that scaled the pilot architecture across additional CNS and non-CNS therapeutic areas enterprise-wide.
Validated against clinician-adjudicated gold-standard cohorts at 75% precision and 70% recall—compressing cohort annotation from months to minutes, saving ~$500K in external clinical labeling costs, and accelerating R&D target validation cycles ~2.5×.
Private Equity Portfolio & Industrial AI Strategy, Disruption Diagnostics & Internal Builds
AI Advisor — "The Builder" at Attian AI (attianai.com) · Private Equity Portfolio Companies, Search-Backed Operators, & Fortune 500 Industrials
Operational Challenge
42% of companies abandoned most AI initiatives in 2025 (S&P Global). Private equity operating partners and industrial CEOs face a chasm between boardroom AI mandates and shop-floor reality: off-the-shelf SaaS tools fail on bespoke ERP/MES workflows, while internal pilots stall without governance or unit-economic ROI.
Implementation Architecture
- Deliver AI Strategy & Opportunity Diagnostics for PE funds, search-backed companies, and industrial operators: workflow mapping, build-vs-buy due diligence, competitive AI disruption scoring (across 6 dimensions), and prioritized 90-day implementation roadmaps.
- Architect and build internal production AI systems for portfolio companies—including custom order management & customer-service triage agents, predictive warehouse labor planning, automated regulatory compliance intelligence, and real-time ERP/MES data integrations.
- Lead executive education and technical upskilling so internal engineering and operations teams can operate, evaluate, and extend deployed agents independently.
Every portco engagement is anchored in baseline unit economics, data defensibility, and measurable EBITDA impact—delivering outcomes such as $1M projected annual warehouse labor savings, 50% faster engineering delivery, and 80% YoY revenue acceleration across advised clients.
Krava Core: Verifiable Hardware TEE & Zero-Knowledge Private AI Inference
Inventor / Entrepreneur-in-Residence (Sold to Customer) · Krava.io — Privacy-Critical Enterprise AI, Regulated Manufacturing & Confidential Compute
Operational Challenge
Industrial manufacturers, defense-adjacent suppliers, utilities, and healthcare enterprises cannot send proprietary CAD drawings, BOM pricing, grid topologies, or patient records to multi-tenant cloud LLM relays where TLS terminates at the provider’s API gateway and plaintext prompts are exposed to infrastructure operators.
Implementation Architecture
- Invented and architected Krava Core (krava.io), a verifiable confidential inference runtime where encryption reaches all the way into hardware-attested Trusted Execution Environments (TEEs).
- Implemented client-side cryptographic attestation verification and Encrypted HTTP Body Protocol (EHBP) request sealing (`@krava/sdk` in TypeScript and `krava` in Python) so the Krava relay handles authentication, routing, and prepaid balance reservation over opaque ciphertext only.
- Designed zero-knowledge usage metering that settles token consumption and audit receipts without storing prompts, completions, or conversation state.
Fail-closed cryptographic verification: if enclave measurement hashes, attestation quotes, or ephemeral key bindings deviate from the caller’s pinned policy (`strict` pin mode), the client SDK refuses to transmit a single byte of plaintext. Technology was successfully commercialized and sold to an enterprise customer.
AI EVALUATION, GOVERNANCE & COMMERCIAL DUE DILIGENCE
Why 42% of enterprise AI pilots are abandoned—and the four evaluation gates I enforce to take industrial & PE-backed systems into production.
Self-Serve vs. Enterprise Unit-Economics & Margin-Tail Due Diligence
Before opening self-serve industrial onboarding at Achilles, I audited 1,781 production & staging LLM cost events across chat queries, CSV/Excel tabular parsing, KSeF XML invoice attribution, and PDF engineering manual ingestion. The telemetry proved that 94% of total LLM spend came from non-chat document ingestion—meaning standard SaaS “message caps” fail to protect gross margins without fail-closed page caps and pre-execution credit reservation.
1. Reproduce “Model 0” & Signed Failure Labels
Before claiming ML uplift, I codify the client’s existing spreadsheet or heuristic (“Model 0”) to within ±0.5% and formalize a signed target-variable definition separating true in-service faults from proactive replacements.
2. Deterministic Risk Math vs. Grounded LLMs
All failure probabilities, dollar consequences (PoF × CoF), and reorder quantities are computed by validated survival or demand models (Croston/TSB, conformal intervals). LLMs synthesize grounded citations with zero invented numbers.
3. Outcome-Gated SOW Acceptance
Every pilot-to-production rollout is structured around measurable operational gates: 80–90% BOM line auto-proposal, ≥60% KSeF XML invoice auto-verification, and 95% bootstrap CI improvement over baseline risk rankings.
4. UCM & Confidential Compute Boundaries
Through Unified Context Management (UCM) and Krava Core TEEs, enterprise agents track source freshness, revalidate stale claims, and execute inside hardware-attested enclaves or air-gapped on-prem GPU clusters.
UNIFIED CONTEXT MODEL (UCM) & INDUSTRIAL AI RESEARCH
UCM — Unified Context Model
Probabilistic impact analysis for code changes and enterprise knowledge graphs—a 3-layer architecture (Calibrated Trust, Change Awareness, and Targeted Revalidation) that answers “I changed this schema or function, what downstream agents or pipelines might break, and how confident are you?” Shipped as open-source Rust crates.
CDC: The Boring Technology That Actually Works
Why transaction-log Change Data Capture (Debezium, Oracle LogMiner, Kafka) beats brittle nightly batch ETL and achieves a 4.2-minute latency floor across ERP, MES, and warehouse operations.
Battery Passport 2027: The Traceability Mandate
How EU Regulation 2023/1542 turns mine-to-module carbon, recycled-content, and cell state-of-health traceability into an urgent ERP/MES data architecture requirement by February 18, 2027.
Unified Context Management (UCM) for Agentic AI
A 3-layer reference architecture—Calibrated Trust, Change Awareness, and Targeted Revalidation—adapted from Google Knowledge Vault and NVIDIA predict-time selection for enterprise agent reliability.