TrueVectors

Founder ​

Radhakrishnan Poomari — a founder-operator and senior AI-systems architect with 20+ years building economically-grounded, production AI in domains where accuracy and consequence matter. Named inventor on three US patent applications, with peer-reviewed and sole-authored research behind the work. My focus has always been one question: how do we find the problems that truly matter, understand them deeply, and build systems that create measurable value in the real world?

I started in industrial AI, where failure is physical and expensive. At GE Global Research I pioneered the company's first deep-learning predictive-maintenance deployment — in production across a fleet of 4,000+ turbines, saving on the order of $2M+ per critical turbine event by catching anomalies before they became catastrophic failures. That environment shaped how I think about AI permanently: it must be grounded in evidence, economics, operational reality, and trust. AI isn't valuable because it's sophisticated — it's valuable when it helps people make better decisions, prevent costly failures, and act with confidence. I've carried that discipline through every domain since, and it's the direct lineage behind the work I do now on the reliability of AI systems.

What I'm building: TrueVectors ​

TrueVectors is my independent research project on Agent Behavioral Reliability — built and published on my own time, unaffiliated with my employer — and it's the hardest, most consequential system I've built. AI agents now produce more work than any team can re-check, so the question has changed from "does this look right?" to "did the agent actually do what it claimed — and does the result hold?" TrueVectors answers that with proof, not opinion.

Most "AI checks AI" tools are fully probabilistic: ask twice, get two verdicts, with the rules hidden inside a model's weights. TrueVectors is neuro-symbolic — the model proposes, the math disposes. An LLM does what it is uniquely good at (recognizing intent and extracting an agent's claims against a closed, version-controlled taxonomy of failure modes); a deterministic engine supplies ground truth and can overrule the model outright. Every finding is stamped with where its verdict came from — deterministic, corroborated, refuted, counterexample, property_checked, or bare llm — so a reviewer always knows what is math and what is judgment.

The depth is in how it reaches a verdict. Wherever a question can be settled mechanically, TrueVectors settles it that way — examining the output and the agent's own recorded actions directly, rather than asking a model for an opinion. When a claim needs harder proof, an optional, isolated check searches for a concrete case that breaks it, or confirms it holds across every case it can reach. The same engine reads an agent run's own declared limits — which tools it was allowed to use, what it was allowed to spend — and checks whether it stayed inside them. The result isn't a verdict that drifts from one run to the next; it's one you can reproduce.

Why it matters: it is the accountability layer the AI economy is missing. Organizations are spending billions on AI agents and cannot answer three basic questions — what did it cost, what did it actually do, and will it still work next quarter. TrueVectors turns those into evidence — on real, public AI work, with every verdict reproducible from the cited source (see the audits). The product is local-first and sends nothing off your machine. As AI models become interchangeable, what stays defensible is the part that doesn't guess — verdicts grounded in what the agent actually produced and did, plus the growing record of results already verified.

Track record ​

GE Global Research (2007–2018) — Senior Staff Data Scientist. Pioneered GE's first deep-learning predictive-maintenance deployment ($2M+ saved per critical turbine event across a 4,000+ unit fleet, in production — not a proof-of-concept). Architected digital twins for gas and steam turbines over massive time-series data, and commercialized R&D algorithms into Analytics-as-a-Service on Predix across power, aviation, and healthcare. Led the SmartSignal M&A technical due diligence, advised NBCU Peacock Equity on AI-startup investment (direct VC-evaluation exposure), and earned consecutive GE Management Awards (2013–2014).

Xformics (2018–present) — Co-founder & Chief Data Scientist. Built the AI & Data Sciences practice to ~20 engineers and scientists, leading product strategy and all technical sales. Built "The Optimizer," a causal-ML promotion-optimization platform that separates true incremental lift from baseline, cannibalization, and halo effects — enterprise-validated by Fortune 500 retailers, including a projected 2400 bps promotional-yield improvement for a pilot category deployed in production. Also built Providence / Revelation (enterprise knowledge-discovery and operations intelligence); a credit-risk model for an international bank (hybrid ML + LLM over 25+ macro indicators, 30-day-ahead default prediction); anomaly detection that cut mean-time-to-resolution by ~50%; and domain-specific LLMs with a ~35% inference-cost reduction alongside explainability and compliance.

Autonomous systems research (~3 years). Nearly three years in autonomous systems — autonomous robots for industrial inspection and Autonomous Underwater Vehicles (AUVs) — where reliability is not optional: closed-loop perception and control, operation in harsh environments, and behavior that must be verified when ground truth is hard to recover after the fact.

Background ​

  • 20+ years in AI, data science, and enterprise analytics
  • Pioneered fleet-scale deep learning for predictive maintenance and reliability intelligence; architected digital twins for complex industrial assets
  • Co-founder and Chief Data Scientist — built enterprise AI platforms (Providence, Revelation) and causal-ML systems for demand and promotion optimization
  • Named inventor on three US patent applications (filed at GE and Xformics) — including knowledge extraction and a system-of-systems control for performance and risk outcomes, the same problem family I now work on for AI agents — plus 40 patent publications authored as an IP consultant for IBM, Motorola, and other Fortune 500s
  • Peer-reviewed and sole-authored research — contributing author in the Journal of the Optical Society of America A (2007); sole author of a COVID-19 incidence model on medRxiv (2020)
  • M.S. Computer Systems Engineering, specialized in Intelligent Systems Design (Sweden); GE Certified Six Sigma Green Belt; GE Certified Senior IP Professional
  • Autonomous systems research (~3 years) — industrial inspection robots and Autonomous Underwater Vehicles (AUVs)
  • Domains: Retail, BFSI, Healthcare, Pharma, Manufacturing, Energy, Industrial Operations, Autonomous Systems
  • Prior deployments: Aviation safety, oil & gas, petrochemical refining
  • Full lifecycle: from problem identification, ROI estimation & justification, data evaluation, architecture design, model build and validation, through production deployment and operational adoption
  • Location: Greater Toronto Area, Canada

What I believe ​

  • AI is a tool, not a default. Not every task benefits from AI. The discipline is knowing when to use it and when not to.
  • Depth beats speed. The AI economy rewards velocity. The real economy rewards judgment. I'm interested in the latter.
  • The best problems are found, not assigned. High-value problems rarely announce themselves. Finding them before they become obvious — by combining signals others read in isolation — is its own discipline, and it's the one I trust most.
  • Accountability is the missing layer. Billions are being spent by organizations that cannot answer three basic questions: what did it cost, what did it actually do, and will it still work next quarter. The next generation of AI — autonomous agents and decision systems — must be judged by whether it understands intent, preserves constraints, and behaves reliably, not just by whether it produces an answer.
  • Open thinking creates trust. I publish my frameworks, research, and analysis openly — because credibility is built by showing your work, not by gatekeeping it.

Connect ​