AI strategy and roadmap
Which problems are worth a language model, which are not, and what the winners cost to run. A scoped plan with evals before any code.
Generative AI consultant / ex-IBM / ex-Director of Engineering, YC W20
Latest essay No. 63 / 08 Oct 2026Production AI that survives contact with real users.
I help founders and engineering leaders take LLM features from demo to dependable: agents, retrieval, integrations, and the evals and guardrails that keep them honest in front of customers.
30 minutes, booked and paid through Razorpay. Not ready for a call? Email hire@viralruparel.com, reply usually within 24 hours.
8+
years shipping software
20+
engineers led as Director of Engineering
63
essays on agent engineering
24h
typical reply to a new enquiry
Built with teams backed by
- Y Combinator
- Andreessen Horowitz
- Bain Capital Ventures
- FirstMark
- BoxGroup
- Hawke Ventures
- Abstract Ventures
- Shopify Ventures
- Blockchain Capital
01
What I do
Generative AI consulting for production LLM and agent systems. Six things, done end to end, and every engagement starts with an eval set so we both know what better means.
How an engagement startsLLM integration
Model calls wired into your product with structured outputs, fallbacks across providers, caching, and cost controls that hold at scale.
RAG systems
Retrieval that answers from your own documents: chunking, hybrid search, citations, and a regression set that catches drift.
Agents and automation
Tool-using agents with authorization gates, idempotent side effects, clean retries, and an audit trail you can defend.
Full-stack product engineering
TypeScript, React, Node and Postgres. The product around the model is what makes an AI feature usable.
Architecture and technical leadership
Fractional leadership: system design, hiring, and the engineering habits that let a small team ship every week.
02
Selected work
What it was, what was broken, what changed. Ask for the long version of any of these on the call.
LLM file-to-JSON standardiser
GPT-4ClaudeLangChainOCRTypeScript
HR teams received payroll and employee data as CSV, Excel and PDF from dozens of sources and spent hours standardising it by hand. The platform accepts any format and returns structured JSON.
Outcome
85%
less processing time, and manual formatting errors gone across several HR systems.
Workflow automation platform
MERNAWSMicroservicesTeam lead
Businesses needed deep integrations between their SaaS tools and had nobody to build them. Joined as a founding engineer, left as Director of Engineering.
Outcome
99.9%
uptime serving thousands of customers, with the team grown from 2 to more than 20 engineers.
Watson-powered SAP support chatbot
IBM WatsonSAPRPALive chat
SAP support queries were burying the IT team and responses were slow. An enterprise chatbot wired into SAP, RPA and live chat took the first pass.
Outcome
70%
fewer support tickets, with 85% of queries resolved correctly by the bot.
Token compensation engine
Smart contractsNode.jsReactPostgreSQL
Blockchain companies were still paying people through opaque, manual, error-prone spreadsheets. Built automated token grants and payroll on smart contracts with AI-driven performance metrics.
Outcome
Transparent token distribution for fintech clients, auditable by the people being paid.
Solar PV and battery control platform
IoTMQTTRabbitMQInfluxDBTimescaleDB
Distributed solar and battery storage installations were monitored by hand and could not optimise energy use in real time. One platform now manages them all.
Outcome
Automated energy management across multiple installations in Europe.
03
Writing
Stop Hand-Tuning Prompts. Let Your Eval Set Do It.
Most teams still tune prompts by hand, one guess at a time, with no memory of what they already tried and no proof the new version is better. Automated prompt optimization turns that into a loop: a reflection model reads your eval failures, rewrites the prompt, and keeps only the versions that score higher.Your Agent Writes Code and Runs It Where Your Secrets Live
The feature that makes a code-writing agent useful, running code it generated a second ago, is also the widest hole in your stack. That code is written from input you do not control, and most teams run it in the same process or a plain container that can read every secret and call any host.Your Agent Logs Everything and Can Prove Nothing
A trace with forty-seven agent steps shows an auditor that something happened. It does not show them what the agent decided, why it decided it, or who was accountable when it acted.Your Agent Failed at Step 7. The Bug Was at Step 3.
When a multi-step agent produces a confidently wrong answer, the step that threw the error is almost never the step that caused it. The real culprit is an earlier decision whose bad output only surfaced downstream, so you patch the symptom and the failure comes back next week.Guides
All guides04
About
Enterprise discipline from IBM. Startup speed from a YC founding team. Eight years of putting both into systems that ship.
I am a Generative AI consultant and senior software engineer. I have spent more than eight years delivering full-stack products, designing the architecture under them, and lately wiring language models into platforms that have to work at scale.
At a YC W20 startup I joined as a founding engineer and left as Director of Engineering, leading a 20-person team through platform architecture, product delivery, and the unglamorous engineering practices that keep a product up.
Earlier, at IBM, I built Watson-based support automation that resolved 85% of queries and cut ticket volume by 70%. Since then I have led token-based compensation infrastructure at a blockchain fintech and energy control platforms for IoT fleets in Europe.
Today I help companies implement production-grade AI that creates business value, and I write about what breaks when agents meet reality.
- Now
- Independent consultant
- Before
- IBM, YC W20 startup, Web3 fintech
Stack
- Generative AI and LLMs
- OpenAI GPT-4, Anthropic Claude, LangChain, AI agents, prompt engineering, RAG systems
- AI and ML platforms
- IBM Watson, Hugging Face, TensorFlow, PyTorch, MLflow, vector databases
- Frontend
- React, Next.js, Tailwind CSS, Angular, Vue.js, Chart.js, Ant Design, shadcn/ui, MUI
- Backend and full-stack
- Node.js, Express, Prisma, T3 stack, REST APIs, Swift, MERN
- Databases and storage
- PostgreSQL, MongoDB, TimescaleDB, InfluxDB, vector DBs, database design
- Cloud and DevOps
- AWS, DigitalOcean, Kubernetes, Docker, Git, CI/CD, serverless
- IoT and real-time
- MQTT, RabbitMQ, IoT system design, real-time data, edge computing, edge AI
- Mobile and Web3
- React Native, iOS, Web3, smart contracts, blockchain, dApps
05
How it starts
A call, a scoped plan with an eval set attached, then a build measured against that set.
A 30-minute call
Tell me what you are building and where it is stuck. I will say plainly whether I can help, and what it would take. If I am not the right person, I will say that too.
A scoped plan with an eval set
Which problems are worth a language model, which are not, and what the winners cost to run. The eval set is written before any code, so we both know what better means.
Built against that eval set
Shipped in slices and measured against the same set, so better is a number we agreed on, not an opinion.
- Project
- Scoped delivery, from an MVP to an enterprise system. Scoped on the call.
- Retainer
- Ongoing architecture and technical leadership for a team that already ships. Scoped on the call.
- Reply
- Usually within 24 hours, from hire@viralruparel.com.
06
Start with a 30-minute call.
Tell me what you are building and where it is stuck. I will say plainly whether I can help, and what it would take.
- Response time
- Usually within 24 hours
- Engagements
- Projects and retainers
Ready to talkBook a 30-minute call
30 minutes, booked and paid through Razorpay. A receipt and a reply within 24 hours follow by email.
Not sure yet? Write first
Free, no commitment. Tell me what you are building and I will reply with whether I can help and what the next step would be.