# Nilay Baranwal

> Forward Deployed Engineer and AI Engineer, Prague, Czech Republic. Full stack and AI engineer in Prague. I sit with the people who will use the system, build the agent, and keep it running in production: MCP servers, agent harnesses, evals and the integrations behind them.

Seven years of turning vague problems into working systems and keeping them running in production. At Barclays I sat with traders to build the pre-trade dashboard they use daily, and now I'm building Hive Mind, the MCP server that connects AI agents to the bank's pre-trade systems.
On the side I've built four AI products from zero and deployed each into a real operating environment: Prague nightclubs on event night, marketing teams inside their own analytics stacks, and short-term rentals run by agents.

## Pages
- [Home](https://nilaybaranwal.com/): who I am, and an agent you can ask about my work
- [Work](https://nilaybaranwal.com/work): where I've deployed, what shipped, outcomes, open source
- [Lab](https://nilaybaranwal.com/lab): interactive experiments that run in the browser
- [Writing](https://nilaybaranwal.com/writing): articles and hands-on guides
- [Gallery](https://nilaybaranwal.com/gallery): studies after the New York street art the site is styled on
- [About](https://nilaybaranwal.com/about): background, experience and toolbox
- [Resume](https://nilaybaranwal.com/resume): the full resume, with a PDF download
- [Contact](https://nilaybaranwal.com/contact): email, LinkedIn, GitHub

## Deployments
- [Hive Mind](https://nilaybaranwal.com/work#hive-mind), Barclays, Markets Pre-Trade (embedded with the trading desk), Barclays, 2021 to now: Every desk that wanted an AI assistant was building its own bespoke integration into pre-trade tooling and market data, with its own credentials and no shared audit trail.
- [Pre-trade dashboard](https://nilaybaranwal.com/work#pre-trade-dashboard), Barclays trading desk, Barclays, 2021 to now: Traders were making pre-trade decisions across scattered tools, and a market-data bottleneck slowed them mid-session. Outcome: 40% faster decisions, $2M a year in cost savings, API responses 60% faster.
- [Social Line](https://nilaybaranwal.com/work#social-line) (live: https://www.socialline.app), Marketing teams, inside their own AI agents, 2026 to now: Brands miss the mentions that matter because they arrive as an unranked stream across seven platforms.
- [Partymaker](https://nilaybaranwal.com/work#partymaker) (live: https://partymaker.app), Prague nightclub doors, on event night, 2025 to now: Promoters running 500 to 3,000 capacity events juggle ticketing, guestlists, door staff and cost splitting across spreadsheets and chats. Outcome: Live with Prague venues and labels including Ankali and Harmony Rec.
- [Orvio.ai](https://nilaybaranwal.com/work#orvio) (live: https://orvio.ai), Customers' GA4, Search Console and CRM, 2025 to 2026: Businesses couldn't tell whether ChatGPT, Gemini, Perplexity or Google AI Overviews cited them, or whether it brought in pipeline.
- [Agentic property management](https://nilaybaranwal.com/work#agentic-rentals), Short-term rentals, Prague pilot, 2026 to now: Traditional property managers charge a large cut for messaging, reviews, cleaning dispatch and reporting that agents can do.

## Writing
- [AI agents now outwork humans at Anthropic. Junior engineers are feeling it first.](https://www.linkedin.com/pulse/ai-agents-now-outwork-humans-anthropic-junior-feeling-nilay-baranwal-26ubf/) (2026-10-17): What Anthropic's own workforce report says about hiring, AI usage and output.
- [Anthropic replaced git. Here's why you probably shouldn't.](https://www.linkedin.com/pulse/anthropic-replaced-git-heres-why-you-probably-nilay-baranwal-qtubf/) (2026-10-14): Reading past the headline of Anthropic's post on redesigning source control.
- [Recruiters can now interview your AI twin. Here's what it actually checks.](https://www.linkedin.com/pulse/recruiters-can-now-interview-your-ai-twin-heres-what-checks-baranwal-yvoaf/) (2026-10-13): A look at TwinCafe, the platform that builds an AI twin from your résumé and work.
- [Everyone's building a Jev rival. The independent numbers tell a different story.](https://www.linkedin.com/pulse/everyones-building-jev-rival-independent-numbers-tell-nilay-baranwal-da8zf/) (2026-10-06): OpenAI, Amazon and Cloudflare launched decision models in three weeks. The benchmarks, checked.
- [Jev vs OpenAI Decisions API: which decision model should your agent use?](https://www.linkedin.com/pulse/jev-vs-openai-decisions-api-which-decision-model-should-baranwal-qlrbf/) (2026-10-02): Comparing the two for classification, routing and agent decisions.
- [Top 10 Jev use cases in agentic apps](https://www.linkedin.com/pulse/top-10-jev-use-cases-agentic-apps-nilay-baranwal-b6rwf/) (2026-09-29): Where a decision model that returns typed answers fits inside an agent.
- [Jev vs Laya: decision models](https://www.linkedin.com/pulse/jev-vs-laya-decision-models-nilay-baranwal-8x1sf/) (2026-09-23): Two launches of the same new kind of model, hosted and open-weight, three days apart.
- [LLMs or Jev: where each one actually fits](https://www.linkedin.com/pulse/llms-jev-where-each-one-actually-fits-nilay-baranwal-cdnsf/) (2026-09-22): Jev and Structured Outputs both promise a trustworthy response shape. They differ underneath.
- [Everyone's talking about Jev. Here's a closer look at what the data shows.](https://www.linkedin.com/pulse/everyones-talking-jev-heres-closer-look-what-data-shows-baranwal-h3zjf/) (2026-09-22): What TypeSafe AI's "System One model" launch data does and doesn't show.

## Lab (interactive experiments)
- [Live agent with an approval gate](https://nilaybaranwal.com/lab/live-agent): Claude answers questions about my work by calling tools over this site, and stops for your approval before it acts.
- [AI engineer interview, beginner path](https://nilaybaranwal.com/lab/ai-interview): Four rounds: core-concept flashcards, system design (including an MCP server), coding you can run, and STAR behavioural answers.
- [Behavioural interview coach](https://nilaybaranwal.com/lab/behavioural): A question bank with what interviewers are really asking, a story bank, timed practice, conflict resolution with the LEAD method, and a weak-vs-strong example.
- [Interview drill, experienced](https://nilaybaranwal.com/lab/interview): Timed FDE and AI engineer questions, the rubric a strong answer covers, and optional feedback from Claude.
- [AI adoption planner](https://nilaybaranwal.com/lab/adoption): Pick a customer, size the time an agent gives back, prepare for the objections, and copy a pilot proposal.
- [Discovery call simulator](https://nilaybaranwal.com/lab/discovery): A first call with a customer who says "we need AI". Choose your questions; the notes you leave with show how well you listened.
- [Scope a vague problem](https://nilaybaranwal.com/lab/scope): Pick a fuzzy request, answer the discovery questions I'd ask on site, and get a scoped first week.
- [MCP analyzer](https://nilaybaranwal.com/lab/mcp-analyzer): Connect to any MCP server, this site's included, and get a graded list of what's wrong with its tools and how to fix it.
- [MCP inspector](https://nilaybaranwal.com/lab/mcp-inspector): Browse a sample MCP server's tools, schemas and annotations, call them, and read the JSON-RPC.
- [Tool gate playground](https://nilaybaranwal.com/lab/tool-gate): Write tool definitions, set the harness policy, and watch which agent calls run, wait or get denied.
- [Prompt-injection spotter](https://nilaybaranwal.com/lab/injection): Documents with hidden instructions meet an agent. Switch guardrails on and off and see which attacks land.
- [RAG chunking visualizer](https://nilaybaranwal.com/lab/rag): Change chunk size and overlap, ask a question, and see which chunks retrieval would hand the model.
- [Eval bench](https://nilaybaranwal.com/lab/evals): Score two versions of an agent against a golden set and find the regressions before your users do.
- [Agent trace debugger](https://nilaybaranwal.com/lab/traces): Four recorded agent runs, each with a classic failure. Read the trace, diagnose it, and see the fix.
- [Prompt caching simulator](https://nilaybaranwal.com/lab/caching): Reorder the parts of a request and watch cache hits and cost change over a conversation.
- [Context window budget](https://nilaybaranwal.com/lab/context): See how system prompt, tools, history and documents fill a context window, and what compaction buys you.

## Open source
- textagent (Alpha, https://github.com/alphasr/mob-agent): Build AI agents people text instead of apps they install. Write the agent once and it runs over iMessage, WhatsApp, Telegram and email.
- Help Engineer (In design): Know what your AI changed, and what it didn't. Test evidence on every PR that the code follows your team's rules and existing behaviour still holds.

## Experience
- Barclays, Senior Full Stack Software Engineer, May 2021 to now. Embedded with the Markets pre-trade desk. Primary technical contact for traders and desk heads: demos, discovery, requirements and release plans.
- TEDxKanke, Lead Full Stack Developer (remote), May 2019 to Feb 2022. Sole technical partner to a non-technical organising team. Scoped and launched the event platform; registrations grew 250%.
- Nationale-Nederlanden, Junior Software Engineer, May to Aug 2018. Eight PeopleCode modules for an HR system serving 3,000+ employees, with requirements gathered from regional HR teams. Reports 40% faster.
- Commerzbank, Junior Software Engineer, Jan to Apr 2018. Spring Boot work on banking applications serving 100,000+ customers daily.

## MCP server
- https://nilaybaranwal.com/mcp (Streamable HTTP, read-only, no auth): search_site, get_deployment, list_writing, list_open_source, get_principles

## Contact
- Email: nilaybaranwal@gmail.com
- LinkedIn: https://www.linkedin.com/in/nilay-baranwal
- GitHub: https://github.com/alphasr