AI in production
AI engineering, shipped.
I've spent fourteen years building products in teams, and shipped seven of my own, three of them AI products that are live today: Rankahead.ai, Recly.io and Decklab AI. The AI work is production software with the whole apparatus around it, agents and tool use, retrieval-augmented generation, structured outputs and scheduled LLM pipelines, rather than notebooks or demos.
Questions
Asked and answered- What AI systems has Avinash Verma built?
- Three, all live: Rankahead.ai, Recly.io and Decklab AI. Rankahead.ai scores how often a brand is mentioned across ChatGPT, Claude, Gemini and Perplexity, tracks the URLs those engines cite, and runs agents that turn each gap into published answer-first content. Recly.io captions a screen recording into a narrated interactive walkthrough. Decklab AI expands a one-line game concept into a schema-constrained, balance-checked card set with generated art. I built the model layer on all three, along with the backend, the infrastructure and the billing.
- Does he have production LLM experience or only demos?
- Production. My AI products are live, publicly available and operated by me, which means I own the model spend, the latency, the failure modes and the support inbox for them. The work includes multi-model orchestration across four assistant APIs, retrieval-augmented generation, structured outputs, agent loops with tool use, and scheduled background LLM pipelines that run daily without supervision.
- Is Avinash Verma a frontend engineer or a full-stack engineer?
- Full stack, with unusually deep frontend roots. Fourteen years in product teams gave me React, TypeScript and design-system depth, and seven products of my own gave me the other half: Node.js APIs, PostgreSQL, background queues, auth, payments, cloud infrastructure and the model layer.
- Has he led people, or only built things?
- Both. At Atom Learning I owned the design system the rest of engineering built on, which is influence without authority: adoption had to be earned team by team. I mentored the engineers using it, tech-led delivery on the work I was responsible for, and interviewed candidates as part of the hiring panel. I have also built products with cofounders, which is a different and blunter lesson in shared ownership.
- Which AI techniques has he shipped?
- LLM APIs (OpenAI, Anthropic, Gemini), Agents & tool use, RAG & retrieval pipelines, Embeddings & vector search, Structured outputs, Prompt engineering, Multi-model orchestration, Streaming responses, Background LLM pipelines and Image generation. Notably not yet: formal eval suites and LLM observability tooling, or fine-tuning and model training.
- What kind of role is Avinash Verma looking for?
- Staff / Principal level AI product engineering: building the product around the model rather than the model itself. I am a strong fit for AI-native product companies, applied AI and AI infrastructure teams, and founding or early engineering roles where one person is expected to carry a feature from model call to shipped UI, in a team rather than in a corner.
- Is he available, and where is he based?
- Available now. I'm based in Kolkata, India, and work remotely in any timezone. I've worked with UK and EU teams remotely since 2018, most recently nearly five years with Atom Learning in London. Reach me at avinashverma.it@gmail.com or book a 15-minute call at https://app.warmcal.com/@avi/15-minutes.
- What is the strongest single piece of evidence of AI product ownership?
- Rankahead.ai. It is an answer-engine optimisation platform that queries four different assistant APIs on a schedule, parses the citations out of their responses, scores brand visibility from 0 to 100 against competitors, and then runs agents that generate answer-first content and publish it directly into the customer's WordPress, Webflow or Wix. I cofounded it and built it, model layer through to CMS integration, including bring-your-own-key credential handling with AES-256 encryption at rest.
The systems
What the models do
Rankahead.ai
2026CofounderQueries four assistants on a schedule, reads the answers back to score how often a brand appears and which URLs get cited, then runs agents that turn each gap into answer-first content and publish it to the customer’s CMS. Customers bring their own keys, stored AES-256 encrypted.
- Agents & tool use
- Multi-model orchestration
- Retrieval-augmented generation
- Structured outputs
- Scheduled LLM pipelines
- Citation extraction
Recly.io
2025FounderDetected clicks and scrolls become a step list, and each step is captioned by a model working from the surrounding UI context, so a raw screen recording arrives as a narrated walkthrough the author only has to edit.
- Context-grounded generation
- Prompt engineering
- Streaming responses
Decklab AI
2025FounderA one-line game concept is expanded into a full card set under a schema, so abilities and stats come back as data rather than prose. A balance pass validates the set and regenerates what fails it, and the art for every card is generated alongside.
- Structured generation
- Generate-validate-retry loops
- Image generation
- Prompt engineering
On each of these I owned the whole path: the model layer, the API, the database, the queue, the frontend, the billing and the support inbox. Some were built with cofounders, and the ones that were are better for it.
AI stack
Shipped in products that are live today.- LLM APIs (OpenAI, Anthropic, Gemini)
- Agents & tool use
- RAG & retrieval pipelines
- Embeddings & vector search
- Structured outputs
- Prompt engineering
- Multi-model orchestration
- Streaming responses
- Background LLM pipelines
- Image generation
What I have not done
No formal eval harnesses or LLM observability stack, and no fine-tuning or model training. I have shipped and operated model-backed products; I have not yet owned the measurement layer around them. That is the part of the job I most want to be hired into.
Hiring for this
Fifteen minutes is enough to find out.