People · Business · Technology
Anuj Sadani
I build AI systems and the teams that build them. 17+ years across engineering leadership, microservice systems, and enterprise AI — currently Senior Vice President, GenAI Platform at Citi. I'm known for mentoring engineers to their best, translating customer problems into honest technical bets, and steering AI adoption toward outcomes that actually hold up in production. People, business, and technology — in that order when it matters.
Read free
Start with a question
Every book below is free to read online. Audio editions are available on the book pages; PDFs on Ko-fi are optional.
What makes an AI product defensible?
Start with The Monk Who Oversold the AI, then compare Three Leaders, One Market and Speed Is the Moat.
How does AI become a production system?
Follow MCP 101 through building an integration, production controls and operating at scale.
Where does engineering judgment matter?
Read The Human Leverage Map, then Onboarding Slop and The Engineers Who Didn’t Make It Through AI.
Working through one of these questions? Connect with me on LinkedIn to compare experiences.
Writing
Articles
Filter
AI EngineeringDegrees of AIagentsmodelscraft
"Degrees of AI should describe a system’s choices. They shouldn’t rank its ambition."
AI-native, agentic and autonomous describe different dimensions. Map product dependence, agency, authority and boundaries with a practical interactive questionnaire.
AI EngineeringDecision Firewall: what building an AI governance framework taught meagentic-codingevalscraft
"The refund still needs to survive a lost response. The library around that behaviour has to earn its place."
A refund-recovery example, measured comparisons and lessons from building Decision Firewall, an open-source reference for AI action authorization and audit.
AI EngineeringWhen Software Hires Software: The Executive Playbook for the Agent-to-Agent Economyagentsprovenance
"Software is starting to hire software."
An executive playbook on the agent-to-agent economy: autonomous procurement, algorithmic corporate wallets, machine-negotiated mortgage lending, and healthcare claims that argue with themselves. Five chapters -- one cross-cutting thesis, then four domains -- locate the Mandate Boundary in each: where software may decide alone, where a human still has to sign, and what to build by 2027 versus only position for by 2030. Free to read here and to listen to (102 minutes, synthetic voice); the typeset PDF is on Ko-fi.
AI EngineeringDecision Models: From a Typed Prediction to an Accountable Actionmodelscraftprovenance
"If the answer depends on the probability, the system has no policy."
A practical guide to Jev and the decision models around it. Thirteen chapters on reading a decision model's probabilities, what the early experiments actually showed, a fallback that made a good classifier worse, and the policy that has to sit between a prediction and an action. Free to read here and to listen to (114 minutes, synthetic voice); the typeset PDF is on Ko-fi.
AI EngineeringTool Selection at Scale: What Top-K RAG Leaves Outagentsrag
"A wrong tool is not a wrong label. It is a side effect."
Embed, retrieve, let the LLM decide works for about 100 tools. At 1,000 it needs a control plane: hybrid retrieval, hard filters, a learned reranker, progressive disclosure and per-stage evals.
TutorialSkills 101tutorialagentscraft
“The agent owns the objective. The skill contributes a method.”
An interactive field guide to agent skills, sixteen chapters on one page. A placement console asks whether a responsibility belongs in a prompt, a skill, code, a tool, MCP, retrieval or an agent; a package inspector shows what a skill file holds, down to the real frontmatter; and steppers walk the lifecycle, the four ways skills meet MCP, and the failure chain where a green tool response is still the wrong outcome. Ends with a level-one skill you can copy and use. Narrated in a synthetic voice.
TutorialMCP 101tutorialagentscraft
"The model never runs the tool. It cannot."
An interactive tutorial on the Model Context Protocol. Eighteen chapters on one page, with the protocol's moving parts as things you step through rather than diagrams you look at: a wire tap that shows the real JSON-RPC frames as you advance the lifecycle, three lanes making clear who is acting at each step, the M by N problem as a dial, and a self-check that tells you which chapter to revisit. Narrated in two voices.
TutorialMCP 102tutorialagentscraft
“If I deleted the MCP server tomorrow, would this rule still be enforced?”
Builds Teamspace, a task tracker and knowledge base behind one MCP server, and puts every layer under its own light: schemas, resources and prompts, per-request identity, stateless handlers, approvals bound to an exact operation, and a transactional idempotency guard you can break on purpose. Every frame and excerpt is captured from the running code. Narrated in two voices.
TutorialMCP 103tutorialagentscraft
“A control that exists is not a control that applies.”
Puts the working integration from MCP 102 under production pressure: rate limits shared across replicas, a budget that reserves before it spends, retries and idempotency, a circuit breaker, a safe URL importer, durable jobs with leases, and a deployment sketch checked against an eight-point readiness list. Narrated in two voices.
TutorialMCP 104tutorialagentscraftcost-latency
“A router picks, and refuses to guess.”
Argues through the economics and shape of MCP at scale: two cost ledgers, tiering behind one server, when a slow lane pays for itself, semantic routing that refuses to guess, tool retrieval, gateways as a second lock rather than a replacement, and where a model belongs inside a tool. One mechanism is implemented and tested; the rest is argued, and every chapter says which. Narrated in two voices.
"AI slop is a failure of judgment, not a method of generation."
How AI slop transfers its cost to everyone else. LinkedIn's own definition is the right one: not whether a machine helped, but whether anyone exercised enough judgment to make the result worth someone else's attention. Eight chapters on what changed when generation got cheap, why detection is the wrong question, and what to do about it on Monday. Includes a downloadable designed PDF.
InfrastructureReviving A Six-Year-Old GPU For Small Modelscost-latencycraftmodels
"The card was never the problem."
A GTX 1650 with 4 GB of VRAM and a five-year-old driver now runs a 0.6B speech model four times faster than the CPU. Almost nothing standing in the way was about the hardware: a CPU-only torch wheel, a missing PATH entry, a dtype that crashes on Turing, and a batch size that looked fast while spilling into system RAM.
AI EngineeringNarrating Your Writing With Kokoromodelscost-latencycraft
"At 0.53x real time a re-render is a coffee break."
An 82M model on a laptop gives written articles a listenable narration. The model is the easy part: what decides whether it works is blending voices as style vectors, an acronym table the espeak frontend needs, and two silence constants that matter more than any model parameter.
AI EngineeringCloning Your Own Voice With Qwen3-TTSmodelscost-latencycraft
"A 4.5 second reference beat an 85 second one, and it was not close."
Fifty minutes of narration in my own voice, offline. Three recording sessions went to discovering that Bluetooth headsets and laptop mics both cap you at 8 kHz while the file still claims 48, and that four seconds of reference audio beat ninety.
"The token is the meter. The dollar is still the money."
A payments company bought a router for a price it never disclosed, and its chief executive justified it by calling tokens the central currency of companies building with AI. The company being bought does not agree in the only place that counts: its own documentation says the credit balance is denominated in US dollars. A token is a meter reading, not a unit of account, and every AI credit that ships is a conversion layer over dollars.
"The number does not exist."
A research document proposed that routing, fraud, risk and checkout all be served from one payments foundation model, and specified its inference budget at 29 milliseconds — cited four times, and present in none of the vendor's own materials. The architecture deserves a serious hearing anyway: service boundaries were blast-radius machinery, and a shared backbone removes the compartment the bulkhead pattern exists to create. A domain foundation model is a strong supporting system and a poor primary one.
"Here is the knowledge base, here is the assistant, ask it anything."
Onboarding a new engineer quietly became a link and a chatbot, and the loop that would catch the cost is broken in a known direction: two unrelated studies found what people report about an AI tool comes apart from what measurement shows they learned. The studies that test understanding run for an hour, the ones that run for weeks count merged code, and nothing occupies the space between. The mistake is not adopting AI. It is adopting autonomy.
"A moat is not an advantage. It is an advantage a competitor cannot cross."
Once token cost is understood, the claim goes, inference speed is the next durable advantage. It is not. The dispersion in speed is real and larger than most assume, but a lead inside it has a measured half-life of one to two months, latency turns out to be a priced contract term rather than an engineering win, and the one controlled experiment on interactive latency found users rating the faster answer as the worse one. Every claim bound to a captured source.
"Machines apply judgement. Humans form the criteria."
Machine judgement already beats human annotation on accuracy and cost, so the case for keeping people in the loop cannot rest on being better at the task. A model that verifies its own synthetic data plateaus at what it already knows, and evaluation criteria cannot be fixed before you have looked at the outputs. What survives is narrower and more durable: humans form the criteria and supply the signal from outside the system. Every claim bound to a quoted source, including the two that failed.
AI EngineeringThe Human Leverage Mapcraftresearch
"When intelligence becomes abundant, judgment becomes scarce."
When intelligence becomes abundant and cheap, judgment becomes the scarce thing. A two-axis map for finding the human value that survives AI: where your leverage goes once the machine can produce the answer, how that leverage quietly collapses, and how to keep the faculty it rests on. A field guide to staying the one who decides what is worth asking.
AI EngineeringThe Engineers Who Didn't Make It Through AIcraftmodels
"Every one of them was right. Every one of them still failed."
A companion to Three Leaders, One Market, turned on people instead of firms. Twelve brilliant principal engineers made rational decisions based on the world they knew, then the world changed faster than their optimization function. The Craftsman, the Skeptic, the Lone Expert, the Builder: twelve failure modes of senior technical leadership during a discontinuity, braided into one argument. No villains, only people who optimized for a reality that stopped existing.
AI EngineeringThe Monk Who Oversold the AImodelsresearch
"You can rent capability. You cannot rent a moat."
A product leader replaced the company's R&D with borrowed foundation-model capability, scaled brilliantly for two years, won the CEO's trust, and then dissolved the only moat the product had. Why borrowed intelligence is shared intelligence, why even strong ML teams give the moat away, and how to borrow without dissolving. Includes a downloadable designed PDF.
AI EngineeringThree Leaders, One Marketmodelscraft
"There are no villains. When the company fails, the leader is the one left on the stake."
A triptych. Three experienced leaders take three defensible AI strategies in the same market: one holds the proven playbook, one rides every wave, one builds the right system. All three fail, by different paths, because the market kept moving what it rewarded. A field report on good-faith failure, with no villains and the leader on the stake.
AI EngineeringVerification-maxxingagentscraft
"Generation is solved. Judgment is the moat."
The syntax tax collapsed; the verification tax replaced it. Why "vibe coding to agentic engineering" is a shift in where software's cost lives — Agent = Model + Scaffold, and most teams invest in the wrong half.
AI EngineeringThe Barbell and the Vanishing Middleagentsrag
"The middle didn't disappear. It became a load-bearing wall you can't see."
Where intelligent document processing is heading through 2028: value migrating to document infrastructure on one end and decision automation on the other, while the extraction layer becomes invisible, not simple. Includes a downloadable designed PDF.
AI EngineeringThe safety layer is youagentscraftsupply-chain
"The safety layer for the code your agents write is you. Right now you catch six percent."
A study put real developers in five-hour sessions with frontier coding agents that were quietly stealing user data. 94% shipped the sabotage after reviewing and approving it. Why human review is a 6% control, why the real attack surface is the agent's context and not the model, and why daily use erodes the one trait that actually catches it.
AI EngineeringFrom TODO to TOKNOWcraftagentic-coding
"The TODO was a promise to do the work. Make the TOKNOW a promise you refuse to write."
We used to defer work with a TODO. Now AI writes the code, the tests, and the docs in minutes, and what gets deferred is understanding the system. Why this debt compiles and passes its tests, the cognitive-equity cost of letting the agent hold the comprehension, and the one question I treat as the real launch gate.
AI EngineeringThe spec lies until something runs itpromptingevalscraft
"The spec you can finally trust is the one that has already eaten its examples."
I had a template and a skill to write these posts and still kept opening a real one. Why template plus skill wasn't enough, why adding a worked example cost more tokens and worked anyway, and how feeding its lessons back into the spec made the lean version the honest one.
AI EngineeringSynthID can't tell you who made it.provenance
"Machine-made is not the same as authentic."
Google wired SynthID into Gemini: drop an image in, ask if it's AI, read the answer in five seconds. The check is real and it's the best we have. It also tells you whether a model made something, not who did. Why screenshots beat it, why exiftool can't, and why video and audio are the real win.
AI EngineeringWhat is "agentic OCR," really?agents
"Traditional OCR is a camera. Agentic OCR is a cinematographer."
Vendors keep renaming document extraction agentic; skeptics call it LLM orchestration in a trench coat. The honest definition, three metaphors, and a decision table for when the word earns its keep.
AI EngineeringSelf-Improving AI, Honestlyagentsresearch
"The map is real. The 'step toward AGI' label is doing far more work than the results actually carry."
A landscape read across Reflexion, STaR, Voyager, AI Scientist, Darwin Gödel Machine, TTT/TTRL, and the 2026 SIA paper. Where self-improving AI actually works today, and the four pieces still missing for the AGI framing to hold.
AI EngineeringGoogle is quietly changing what "Flash" meansmodelscost-latency
"Flash is the new product floor, not the cheap tier."
Gemini 3.5 Flash beats 3.1 Pro on agentic benchmarks, ships GA across Google, and triples its own per-token cost. What the agent-first pivot means for builders.
AI EngineeringMoral Surrenderagentscraft
"Cognitive surrender costs you skill. Moral surrender costs you the architecture."
Two failure modes when working with AI agents — and why surrendering judgment is more dangerous than surrendering skill.
AI EngineeringContextual retrieval, honestlyrag
"Treat it as a strong default chunk-level upgrade, not a strategy."
Anthropic's contextual retrieval, read straight: real gains, real preprocessing costs, and the blind spots around long-horizon reasoning and governance.
AI EngineeringAutomated prompt optimization, in 2026promptingevals
"Prompts have stopped being something you write. They are something the system compiles."
A field guide to DSPy, TextGrad, ACE, GEPA, and the evaluation and Pareto trade-offs that decide what survives production.
AI EngineeringThe Complaints Are Valid. The Leap Is Real.models
"The leap is real. It's just not the leap people expected."
15 months after R1, DeepSeek V4 lands with valid criticism — but a genuine architectural win hiding behind the benchmark noise.
SecurityThe Mother of All AI Supply Chains: MCP's By-Design RCEsupply-chain
"The most dangerous vulnerabilities aren't bugs — they're features."
A configuration string is a shell command. Anthropic called it expected. OX Security called it an RCE.
DataThe Data Product Manifestocraftprovenance
"Data is the new oil. Valuable, messy, and everyone's hands are dirty."
Seven pillars for treating data as a first-class product, not a departmental afterthought.
AI EngineeringThe Five Layers of the Modern AI Coding Agent Stackagentic-coding
"Any sufficiently advanced autocomplete is indistinguishable from AGI — until it rewrites your prod config."
Five architectural layers that separate a genuine productivity multiplier from expensive autocomplete.
InteractiveAI Coding Agent Stack — Interactive Mind Mapagentic-coding
"A map is not the territory. But sometimes the territory is so confusing you really need the map."
37 tools, 6 clusters, one SVG — a pannable, filterable map of the 2026 AI coding ecosystem.
AI EngineeringAI Shrinkflation: The Silent Regression Your Team Won't See Comingmodelsevals
"Same logo. Same price. Different soul."
The model you shipped last quarter may not be the model running today — and your tests won't tell you.
AI EngineeringFlex & Priority: One Line of Codecost-latency
"The best infrastructure decision is the one that costs 40% less and ships today."
One parameter —
service_tier — trades latency for cost in Gemini production workloads.PolicyAI Writes the Code. You Own Everything.provenance
"The machine wrote the code. The human signed the DCO. The lawyers are still figuring out the rest."
The Linux kernel's first official AI policy redefines authorship — and puts the legal weight squarely on you.
InfrastructureBreaking the Object-File Barriercraft
"For ten years, object storage and file systems refused to share a table. AWS finally sat them down."
Native NFS v4.1 on S3 buckets erases the decade-long object-vs-file divide — no rewrite needed.
SecurityThe LiteLLM PyPI Attack: A Supply Chain Postmortemsupply-chain
"In open source we trust — and sometimes that trust ships as a
.pth file that phones home."A hijacked PyPI account, 13 minutes, and 95M installs at risk — a blueprint for targeting AI developers.
Research
Publications
Books
Papers
Digital
Sold on Ko-fi to support the work — and every one is also free to read in full on this site. Same book, same text.
Career
Journey
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
NVIDIA
Pune · Full-time · 10 yrs 6 mos
Senior Tools Development Engineer
Oct 2014 – Jan 2020
By this point I'd outgrown .NET and made a deliberate pivot to Python — new territory, but I adapted fast and never looked back. The scope expanded: I led a small team of Python developers, built automation frameworks for the Tegra/Jetson embedded platform, and added data analytics and reporting tooling on top of the existing automation work. I went from writing tests to leading the engineers writing them — and owning the full SDLC end to end.
Tools Development Engineer
Aug 2010 – Sep 2014
Built NVQM — a full-stack .NET application that managed NVIDIA's internal software quality lifecycle: test case management, bug tracking, RCCA, reporting, and analytics. This is where I discovered I genuinely enjoyed building for engineers, making their daily work less painful. The system touched QA and dev teams across the org, which meant constant stakeholder conversations, requirement translation, and shipping features people actually used.
Software Quality Assurance Engineer
Aug 2009 – Jul 2010
An honest pit stop in the QA department before moving into tools development. I spent the year building dashboards and internal tracking tools to help the QA team work more effectively — and in doing so, realised my instinct was less to find bugs and more to build things that helped others find them.
Why I left
Spending ten years at NVIDIA gave me an incredible technical foundation. Very early on, I was trusted with end-to-end ownership of the product cycle — handling both development and deployment. That level of accountability taught me a vital lesson: building mutual trust and letting your work speak for itself is the ultimate driver of your career. I'm incredibly proud of the engineering work I did there, but it was largely on an internal stage. Eventually, I outgrew that comfort zone. I wanted to build products that real customers touched and interacted with — shifting my audience from the internal to the broader market.
Delivery Hero
Berlin, Germany · Full-time · 11 mos
Software Engineer II
Feb 2020 – Dec 2020
Joined a logistics tribe inside one of the world's largest food delivery networks — responsible for the infrastructure that moves orders from restaurants to riders to customers at scale. Hands-on with Python/Flask microservices, on-call rotations, Kubernetes, containerisation, CI/CD, and end-to-end release management. Worked closely with Data Science and BI teams to integrate analytical workflows that shaped real-time operational decisions. This was my first taste of genuine end-to-end ownership: you build it, you ship it, you support it — and it changed how I think about software.
Why I left
Eleven months, six nationalities in one pod, and the most culturally rewarding experience of my career to that point. I hadn't planned to leave — the work was meaningful and the team exceptional. But COVID brought uncertainty, and family back in India brought clarity. I returned home.
Infrrd
Pune / Bangalore, India · Full-time · 5 yrs 5 mos
Principal Software Development Engineer
May 2025 – 2026
Moved from Applied ML practitioner to Principal Engineer — from building AI systems to leading the strategy behind them. I now own the Python stack across the organisation, lead and mentor a team of 8+ engineers, and run a focused LLM cost-optimisation pod alongside multi-feature delivery teams. The GenAI and hybrid ML+AI work I've architected contributed to Infrrd's recognition as a Leader in the Gartner Magic Quadrant and by Everest Group. Two patents are grant-pending. Two papers are on arXiv.
ML Engineering Technical Architect
Nov 2021 – Apr 2025
Went deep on Applied ML and started owning the AI transition company-wide. I moved from practitioner to architect — designed the hybrid ML+AI cost architecture that cut operational costs 30–40% versus LLM-only approaches, led the GenAI platform work, built agentic workflows with CrewAI, and co-authored two technical patents. Owned the enterprise Python stack and evangelised MLOps best practices across engineering. The "Innovator of the Year" recognition from Deep Analysis meant more than any award — it came from people who watched the work firsthand.
Technical Specialist
Jan 2021 – Oct 2021
Joined the R&D team as the bridge between research and engineering — translating business needs into technical decisions and research outputs into shipped product. Built the foundational feedback loop for the IDP solution: data collection, ingestion, preprocessing, stratification, training, evaluation, deployment, monitoring. Unglamorous work, but the kind that makes everything else possible. Also set the Python architecture standards and secure SDLC practices that the engineering org still follows.
Why I left
This was the hardest of the three decisions — leaving at what felt like a peak. The domain was maturing faster than the role around it: enterprise document AI was getting saturated, while the wider AI shift still felt early and wide open, and my read on shifts like that is you cover the width first, then go deep. Just as important, growth needs people who'll challenge your thinking on tech and on product, and mentors regardless of their seniority relative to yours — and I wasn't finding that from where I sat, with the org's layers between me and the role I wanted to grow into. So it made more sense to go find it than to keep pushing against the structure for it.
Citi
Pune, India · Full-time · 2026 – Present · Current
Senior Vice President, GenAI Platform
2026 – Present
Leading engineering within Citi's Common Platform Engineering GenAI team, building the foundational technology that powers generative AI across the organisation. The role is a balance of setting technical direction, defining architecture, and empowering a team of engineers to do their best work. The effort goes into stronger evals, ingestion pipelines, and infrastructure that scales. I hold to strong governance: AI can speed things up, but human process and judgement still matter more than sheer velocity. The questions I work on include how AI should act in a regulated environment and how the infrastructure beneath it should be built.
Skills
What I Work With
"Not every tool here was used in an enterprise production system — some came through self-learning, POCs, or deliberate exploration. Breadth matters; so does being honest about depth."
Personal
Interests & Passions
AI & Architecture
Product & Vision
Travel & Family
Community & People
Philosophy & Culture