# Pankaj Rajan > Pankaj Rajan is an agentic AI leader and Head of AI & Engineering at Eightfold AI. He builds enterprise multi-agent systems and long-running agents, and advises on AI transformation strategy — grounded in 15+ years of AI at Walmart, Alexa, and eBay. This is the full text of https://www.pankajrajan.com/. A shorter summary is at https://www.pankajrajan.com/llms.txt. Last updated: 2026-11-02. ## Key facts - Current role: Head of AI & Engineering at Eightfold AI (2025 – Present) - Location: San Francisco Bay Area - Founder: Co-Founder & CPTO of MarkovML; Co-Founder & CEO of Ingenus - Companies: Eightfold AI, MarkovML, Amazon Alexa AI, Ingenus, Walmart Global Tech, Ayasdi AI, eBay, Microsoft, AMD - 15+: Years shipping AI to production - $8M: Raised as co-founder & CPTO, MarkovML - $1B+: GMV enabled by eBay machine translation - Millions: Daily users on AI systems I built - Contact: pankaj@pankajrajan.com; book a 30-minute session at https://calendly.com/pnrajan ## About I have spent 15+ years building AI that real people use — and the last decade leading the teams and companies that build it. At Eightfold AI, I lead AI and engineering for the company’s agentic products: the shared platform agents run on, and the agents, assistants, and analytics built on top of it. I have also architected the agent platform, with guardrails and a knowledge base, and built consumer agents on it for a major frontier AI lab with millions of daily active users. Earlier, as co-founder and CPTO of MarkovML, I raised $8M and built a no-code, data-centric ML SaaS platform for enterprise knowledge work, certified to SOC 2 Type II. Before that, I built evaluation, knowledge (RAG), and safety systems for open-domain dialogue at Amazon Alexa, founded the Query Understanding team behind Walmart.com contextual search, and was the second engineer on eBay’s machine translation engine, the world’s first built for e-commerce. I write about what I have learned building these systems, and I make time each month for a few sessions with founders, leaders, and investors working through the same problems. ## Selected work ### Consumer agents at scale (Frontier AI lab) - Results: Millions of daily active users - Challenge: Ship consumer agents people trust, for millions of daily users. - What was built: Architected the agent platform, with guardrails and a knowledge base, and built several consumer agents on it: shopping, on-demand RAG, and customer service. ### An AI-native product & engineering org (Eightfold AI) - Results: 40%+ faster development · ~60% tickets automated · <1% SLA breaches · migrations weeks → 1 day - Challenge: Slow build cycles. Engineers buried in support, triage, reporting, and migrations. - What was built: Made AI the default way we build: prototype with AI, validate with customers, then ship. Agents now run support, triage, SLA alerts, and exec and MBR reporting, and any team can launch its own from a YAML config. Near-flat headcount; two new initiatives launched. ### One agent platform for every product line (Eightfold AI) - Results: One platform, many agents - Challenge: Every product line needed agents. None could afford its own stack. - What was built: Built one shared platform — runtime, memory, tools, guardrails, evals, tenant isolation — and took agentic products from first code to production on it, including a conversational analyst that turns plain-language workforce questions into answers. ### Open-domain conversational AI at Alexa (Amazon Alexa AI) - Results: User rating up 2× to 3.7/4 · −40%+ latency · −15% cost to serve - Challenge: Raise open-domain quality without raising serving cost. - What was built: Built the evaluation system for conversational agents that gated every model change, an early knowledge service (RAG) that grounded open-domain dialogue, and classifiers for intent detection, anti-hallucination, and safety. ### A no-code, data-centric ML SaaS platform, from zero (MarkovML) - Results: Deployment: 1 week → hours - Challenge: Enterprises had knowledge-work problems and data, but no fast, trustworthy path to a working model. - What was built: Built an end-to-end no-code platform: data quality scored against goals, synthetic data enrichment, bias detection, baseline models, and an evaluation harness — blending classical ML, generative AI, and statistics. Delivered as multi-tenant SaaS, certified SOC 2 Type II. ### AI that moved the needle for defense and EdTech (MarkovML) - Results: FedRAMP requirements met · −80% copy-editing cost · higher throughput, same team - Challenge: Customers needed AI on sensitive data and real workflows, with ROI they could measure. - What was built: Defense: software and a deployment strategy that met FedRAMP requirements, letting teams process sensitive data through no-code workflows in an isolated environment. Publishing: an AI copy-editing workflow that cut cost 80% and raised throughput with the existing workforce. Learning: AI coaches that help students learn in multiple languages. University recruiting: for a large EdTech vendor, data analysis and prediction models that pinpointed the right top-of-funnel population to reach. ### Contextual search for Walmart.com (Walmart Global Tech) - Results: +85% long-tail understanding · ranking launches in days, not months - Challenge: Long-tail queries were misunderstood, and proving out a new ranking algorithm took months of costly A/B tests. - What was built: Founded the Query Understanding team for contextual search — entity extraction, intent classification, and query rewriting — and built an evaluation harness using Bayesian belief networks and click models to test ranking algorithms before they reached A/B tests. ### World’s first machine translation for e-commerce (eBay) - Results: $1B+ in GMV - Challenge: Cross-border buyers could not read listings in their own language. - What was built: As the second engineer on eBay’s machine translation system, worked across the entire pipeline — from conception and model building to evaluation and deployment — opening cross-border marketplaces to buyers in their own language. ## Experience - **Head of AI & Engineering, Eightfold AI** (2025 – Present): Lead AI and engineering across talent management, agentic candidate experiences, career coaching, talent design, and workforce readiness — products used by Fortune 1000 enterprises. Lead the evolution of the agentic platform, combining reasoning models, retrieval, grounding, tool use, predictive ML, and evaluation, and set technical direction across orchestration, enterprise integrations, security, tenant isolation, data residency, and production reliability. - **Co-Founder & CPTO, MarkovML** (2021 – 2025): Built a no-code, end-to-end data-centric ML SaaS platform — data quality against goals, synthetic data enrichment, bias detection, baseline models, and an evaluation harness — combining classical ML, generative AI, and statistics for everyday enterprise knowledge work. Led a 35-person company across engineering, platform, security, and go-to-market. Raised $8M and sold into defense (FedRAMP-compliant deployment) and EdTech (−80% copy-editing cost for publishers, multilingual AI coaching for students), achieved SOC 2 Type II, and presented strategy and risk to the board. Merged into Zig.ai in 2025. - **ML Lead, Conversational AI, Amazon Alexa AI** (2020 – 2021): Built the evaluation system for conversational agents that gated every model change before production, an early knowledge service (RAG) for the open-domain dialogue system, and classifiers for intent detection, anti-hallucination, and safety. Worked on early large language models for open-domain dialogue, alongside response ranking and knowledge graphs. Doubled the user rating to 3.7/4, cut latency by over 40%, and reduced cost to serve by ~15%. - **Co-Founder & CEO, Ingenus** (2018 – 2020): Raised pre-seed funding and launched a multimodal conversational shopping assistant. Wound down when COVID closed retail, returning remaining capital to investors. - **ML Architect, Search Science, Walmart Global Tech** (2015 – 2018): Founded the Query Understanding team behind Walmart.com contextual search, improving long-tail understanding by 85%. Built an evaluation harness using Bayesian belief networks and click models that cut the time to launch new ranking algorithms from months to days and caught costly A/B test mistakes before they shipped. - **ML/NLP Engineer → ML Architect & Lead, Ayasdi AI · eBay · Microsoft · AMD** (2009 – 2015): At Ayasdi AI, architected and built the supervised machine learning platform on topological data analysis — unlocking several million dollars in new revenue in its first year across healthcare, finance, and fraud detection. Before that, the second engineer on eBay’s machine translation engine, the world’s first for e-commerce: worked across the entire pipeline, from conception and model building to evaluation and deployment, adding $1B+ in GMV. Earlier roles at Microsoft and AMD. ## Expertise ### Agentic AI Leadership https://www.pankajrajan.com/expertise/agentic-ai-leadership Most companies now have agent pilots. Few have agents in production that customers trust. As an agentic AI leader, I set the strategy, build the teams, and ship the systems that close that gap — deciding where agents create real leverage, where a workflow or a classical model is the better answer, and what has to be true before an agent acts on its own. At Eightfold AI I lead AI and engineering for the company’s agentic products, from the shared platform to the agents, assistants, and analytics built on it. I have also architected the agent platform and built consumer agents for a major frontier AI lab, and co-founded MarkovML as CPTO. Skills: Agentic AI strategy & roadmap; Agent vs. workflow vs. model: choosing the right tool; Building and leading AI engineering teams; From agent pilot to production; Agentic AI for boards, CEOs & investors; AI-native engineering organizations ### AI Transformation for Knowledge Work https://www.pankajrajan.com/expertise/ai-transformation-for-knowledge-work AI transformation is not a chatbot rollout. In knowledge-work businesses it changes the unit economics: which tasks are automated, which are augmented, and how roles, teams, and workflows are redesigned around digital workers. I advise leaders on AI transformation strategy grounded in what actually ships — sequencing use cases by ROI, risk, and readiness so the investment shows up in the P&L. At Eightfold AI I lead workforce readiness products that help Fortune 1000 enterprises see which tasks will be automated or augmented, how roles will evolve, and where to invest in reskilling and workforce planning. In my own organization I led an AI transformation of how product and engineering work: AI-first prototyping that cut development time by more than 40%, agents that answer ~60% of support tickets automatically, and a no-code agent builder that lets teams launch their own agents. Skills: AI transformation strategy & roadmap; Task-level automation & augmentation mapping; Digital workers alongside human teams; Workforce readiness, reskilling & role redesign; Use-case prioritization by ROI and risk; AI ROI: from pilot to P&L ### Enterprise Agentic AI & Multi-Agent Systems https://www.pankajrajan.com/expertise/enterprise-agentic-ai-and-multi-agent-systems Enterprise agents succeed or fail on the platform beneath them: orchestration, tool access, grounding, permissions, and isolation. I architect agent platforms and multi-agent systems so every product line ships agents on one governed foundation instead of building its own stack. At Eightfold AI I lead the agentic platform that combines reasoning models, retrieval, tool use, predictive ML, and evaluation, with security, tenant isolation, and data residency built in. For a major frontier AI lab I architected the agent platform, with guardrails and a knowledge base, and built consumer agents on it — shopping, on-demand RAG, and customer service — serving millions of daily active users. Skills: Multi-agent orchestration; Agent platforms shared across product lines; MCP, A2A & tool interoperability; Agentic commerce & customer support agents; Agent identity, permissions & guardrails; Tenant isolation, data residency & sovereign AI ### Long-Running & Autonomous Agents https://www.pankajrajan.com/expertise/long-running-and-autonomous-agents The next generation of enterprise agents works for hours, days, and weeks, not single turns. That demands an agent harness with durable state and memory, context engineering and compaction, planning and recovery, human checkpoints at consequential decisions, and evaluation that proves each release is better than the last. I have taken long-horizon agents to production, including an agent that carries candidates from first search to offer letter over weeks of conversation, and assistants with long-term memory that carry context across sessions. Skills: Long-horizon agents that run for days and weeks; Agent harness design: state, memory & compaction; Context engineering for reliable agents; Reasoning-model planning, recovery & escalation; Human-in-the-loop at consequential decisions; Agent evals that gate every release ### Enterprise Search, RAG & Conversational AI https://www.pankajrajan.com/expertise/enterprise-search-rag-and-conversational-ai An agent is only as good as what it can find and understand. I bring two decades of search and language work — query understanding, retrieval, ranking, entity linking, and translation — to the RAG systems, agents, and conversational analytics enterprises are building now. I founded the Query Understanding team behind Walmart.com search (+85% long-tail understanding), was the second engineer on eBay’s machine translation engine (the world’s first for e-commerce, $1B+ in added GMV), and built early LLM and retrieval-augmented conversational systems at Amazon Alexa that doubled user satisfaction. Skills: Hybrid & semantic retrieval for agents; RAG and grounding over enterprise data; Query understanding & learning-to-rank; Knowledge graphs & entity linking; Conversational analytics; Multilingual NLP & machine translation ### Classical & Predictive ML https://www.pankajrajan.com/expertise/classical-and-predictive-ml Not every problem needs a large language model. Classification, ranking, recommendation, and fraud models remain the most reliable and cost-effective answer to many enterprise problems — and the predictive backbone of good agents. At Ayasdi AI I architected the supervised learning platform that unlocked several million dollars in new revenue in its first year across healthcare, finance, and fraud detection. At MarkovML I built a no-code, data-centric ML SaaS platform: data quality scored against goals, synthetic data enrichment, bias detection, baseline models, and an evaluation harness — combining classical ML, generative AI, and statistics to solve everyday knowledge-work problems. Skills: Classification, regression & ranking; Recommendation & personalization; Fraud detection & trust and safety; Predictive ML inside agentic systems; Data-centric AI & human-in-the-loop labeling; Experimentation & offline evaluation ### AI Governance, AgentOps & Cost https://www.pankajrajan.com/expertise/ai-governance-agentops-and-cost As agents move into core workflows, the questions become control, risk, and cost. I help teams put AI governance in place before customers are exposed — evaluation, audit trails, permissions, and release gates — and run agents efficiently by routing across frontier, open-weight, and private models. As co-founder and CPTO of MarkovML, I took a multi-tenant enterprise AI SaaS platform to SOC 2 Type II, with the security, tenant isolation, and evaluation controls enterprise buyers require. Skills: AI governance & responsible AI; AgentOps: tracing, observability & audit trails; LLM evaluation & release governance; Model routing across frontier, open-weight & private models; FinOps for agents: inference cost optimization; SOC 2 Type II enterprise readiness ## Point of view - **A demo is a milestone. Reliability is engineering.** AI lets one person prototype and reach product-market fit faster than ever. But making it work once is a milestone; making it work the thousandth time, for the millionth user, is engineering. The opportunity for leaders is to turn that speed into operational discipline. - **Code is getting cheaper. Understanding is getting more valuable.** AI made output cheap and understanding expensive. When AI writes and AI reviews, understanding fades — until a system misbehaves at midnight. The engineers who understand the system, make the tradeoffs, and own the outcome are where the value is. - **Agents fail on evaluation, not on models.** Most agent programs stall because nobody can say whether this week’s version is better than last week’s. Teams that gate every release on task-level evals ship faster and break less. - **Search is the backbone of every useful agent.** An agent is only as good as what it can find. Query understanding, retrieval, and ranking — disciplines refined over two decades of e-commerce search — decide whether an agent answers or guesses. ## How I evaluate an AI company - **Defensibility:** What remains valuable when frontier models improve next quarter: proprietary data, workflow depth, distribution. - **Architecture:** Whether the system can scale in users, tenants, and use cases without a rewrite — and where the debt sits. - **Evaluation rigor:** How quality is measured and gated before release. Teams without evals are shipping demos. - **Unit economics:** Inference, retrieval, and human-review cost per outcome, and how it trends as usage grows. - **Enterprise readiness:** Security, tenant isolation, compliance, and private deployment — what large customers will demand in procurement. - **Team:** Whether the technical leadership can hire, ship, and make sound build-versus-buy decisions at the next stage. ## Writing ### Making it work once is a milestone. Making it work the thousandth time is engineering. https://www.pankajrajan.com/writing/making-it-work-once-is-a-milestone (2026-10-12) Someone who has built and invested in companies recently asked me “How do you see AI changing how we build software?” My answer: I’m excited and worried. Excited because one person can now prototype, test, and iterate much faster. You can learn faster and perhaps reach product-market fit sooner. Worried because it's most common to think of a convincing demo as a finished product. Making it work once is a milestone. Making it work reliably the thousandth time, for the millionth user, is engineering. For engineering leaders, the opportunity is to turn that speed into operational discipline that creates better products which customers can depend on. AI may help you experiment and reach product-market fit faster. You still have to do engineering work to sustain it. What do you think? ### AI made output cheap, but made understanding expensive. https://www.pankajrajan.com/writing/ai-made-output-cheap-but-made-understanding-expensive (2026-09-24) AI made output cheap, but made understanding expensive. I keep hearing the same story from engineering leaders. Teams ship more code than ever. PR counts are up. Cycle times are down, and they are shipping more. However, under the hood, the story is different. Senior engineers are buried in AI-generated PRs. So they lean on AI to review them. AI writes, AI reviews, humans merge. We regress to convenience, and eventually understanding fades, which bites when systems misbehave. Earlier in my career, senior engineers could easily point to the exact block behind a bug. They wrote it, read it, reviewed it, and debugged it many times. Today, the first thing we do is to paste the stack trace into an LLM and hope 🤞 I do not know where it would end, but I am sure of one thing. As an engineer, the farther you stay from understanding the output, the farther you'll be from the engineer you used to be. ### I can page you. I can’t page Claude. https://www.pankajrajan.com/writing/i-can-page-you (2026-09-17) I can page you. I can’t page Claude. I told this to a software engineer today who was worried about whether software developers will still be relevant in the future. I told him: I’ll give you Claude, and whatever comes next. But I won’t replace you with it because for me you + Claude are 5x more productive. As a team, we can get more done faster. Unfortunately, demos have made it look like software development is disappearing. What’s really happening is that writing code is getting cheaper. But production software still has to be understood, operated, debugged, and improved. And when production breaks at midnight, we need someone who understands the system, can make the tradeoffs, and owns the outcome. That's where value is. Yes, coding is getting cheaper, but understanding is getting more valuable. That's my take at least. ## Links - Website: https://www.pankajrajan.com/ - Full text for AI assistants: https://www.pankajrajan.com/llms-full.txt - LinkedIn: https://www.linkedin.com/in/pankaj-rajan - Google Scholar: https://scholar.google.com/citations?user=-y7FX5wAAAAJ - GitHub: https://github.com/pnrajan - Digital twin (chat): https://www.twincafe.com/about?referral=KNdVJ-dy - Book a 30-minute session: https://calendly.com/pnrajan - Email: pankaj@pankajrajan.com