Agentic AI Engineer @ Sutherland
Qdrant Star | Researcher | Open Source Developer Advocate | Machine Learning Engineer
An Agentic AI Engineer and Qdrant Star who architects RLHF pipelines and multi-stage agentic inference systems to deliver production-ready, highly scalable AI solutions. Currently building tools that let coding agents act on the real web safely.
Ostler is a browser server that AI coding agents drive over MCP (and other agents over A2A). It fills complex web forms in three tool calls, checks its own work, and remembers how in Qdrant.
| What | How |
|---|---|
| One call per form | form_inspect → form_fill → form_submit: typed fields in, intent out, instead of a click and a snapshot per field. |
| Verified, not assumed | Every value is read back from the live page and compared by field type. A page that rewrites a date or picks a different combobox option is reported as a failure. |
| Honest about gaps | Keys that match nothing are UNMATCHED, ties are ambiguous, and a rejected submit returns the page's own validation messages. |
| Hard forms | Dependent dropdowns, multi-step wizards, input masks, comboboxes, shadow DOM and same-origin iframes are handled by an injected runtime. |
| Sees like a user | look returns a screenshot with numbered click targets. Red borders and toasts count as errors. |
| Learns | Profile facts, per-site playbooks and episodes live in Qdrant. Wins and losses per form and per field mapping rank what to reuse next time. Mappings transfer between similar forms on the same site after a redesign. |
| Safe by default | No DevTools port (Chrome is driven over a pipe). Prompt-injection text from pages is fenced as untrusted. Uploads are confined to an allowed directory. Domain allowlist and private-network blocking are available. |
| Hands off CAPTCHAs | reCAPTCHA, hCaptcha, Turnstile and "verify you are human" walls, including search-engine bot checks, are detected and reported as CHALLENGE:. A person solves them via wait_for_human. Ostler never solves or evades them. |
| Robust as an MCP server | Survives a Chrome crash mid-session, leaves no orphan browsers on shutdown or SIGKILL, and runs many sessions side by side. |
Built in Rust with a direct Chrome DevTools Protocol client, 21 MCP tools, and works with Claude Code, Codex, Cursor, VS Code and any other MCP client.
claude mcp add ostler -- ostler serve| Project | Description | Link | Stars | Role |
|---|---|---|---|---|
| Ostler | Rust browser server for AI coding agents (MCP + A2A): verified one-call form filling, Qdrant-backed memory and learning loop, CAPTCHA hand-off instead of evasion. | GitHub |  |
Creator / Core Maintainer |
| t2sql-rlhf | A trustworthy text-to-SQL pipeline with RLHF and vector query caching using Qdrant. | GitHub |  |
Creator / Core Maintainer |
| agentic-rag-skills | 12 production-grade MCP skills built to fix RAG pipeline issues — wrong chunk ordering, token budgets, deduplication. | GitHub |  |
Creator / Core Maintainer |
| Universal Public Data MCP Server | Unifies 20+ real-world APIs into a single interface for AI agents, researchers, and BI dashboards. | Blog | — | Developer |
| Vector Search & RAG Infrastructure | Open-source optimization contributions to vector databases and RAG infrastructure. | GitHub | — | Qdrant Star Maintainer / Contributor |
| visual-evidence-candidate-survival | How many human-annotated relevant pages survive when exact dense retrieval (Qdrant) hands B candidates to a late-interaction reranker. Pinned ViDoRe v3 data, Colab T4 notebook, verified results. | GitHub |  |
Creator |
| payload-audit | Reproduces the dynamic-payload-key finding on a real Qdrant server (3,685 keys indexed one by one cost ~98x the disk of a fixed schema) and ships the audit tool that catches it. | GitHub |  |
Creator |
| agentic-retrieval-bench | Does a small local agent need a vector database? A benchmark with offline tests and CI. | GitHub |  |
Creator |
| qdrant-hybrid-preflight | Check the setup before you trust a hybrid-search result. Packaged CLI, offline tests, CI. | GitHub |  |
Creator |
| local-inference-bench | Real FastEmbed local-inference run versus cited external API pricing, same documents, same collection shape. | GitHub |  |
Creator |
| ask-my-tabs | Your browser tabs, cross-examined: every answer cites its page. | GitHub |  |
Creator |
Pull requests to the projects I use. Status is as of 1 Nov 2026. Full list.
- Published researcher on Computer Vision and NLP at the IEEE ICETITE conference.
- Collaborated on advanced visual AI research alongside Meta Reality Labs during a tenure at CVIT, IIIT Hyderabad.
- Architected production RLHF pipelines from scratch for Text-to-SQL (preference data collection → reward modeling → PPO policy tuning) serving 10K+ daily enterprise transactions on AWS.
- Cut production LLM latency by 60% while maintaining F₁ > 0.95 using multi-stage agentic inference and 4-bit/8-bit quantized model serving.
- Built Ostler, a Rust MCP/A2A browser server with a verified form-filling engine and a Qdrant learning loop, covered by end-to-end tests against real Chrome and a real Qdrant server.
- Qdrant Star — designated for outstanding developer advocacy and contributions to open-source vector search.
- Machine Learning Domain Head — GDSC PICT.
- Core Community Volunteer — TensorFlow User Group (TFUG) Bangalore & AWS User Group Bengaluru.
- Data Science and Machine Learning using Python — iHUB DivyaSampark @ IIT Roorkee.
- AWS Cloud Practitioner Essentials — Amazon Web Services.
Open to collabs, consulting, and open-source contributions.
Ping me — let's build scalable AI systems together.
If my projects helped you, consider fueling more free tools:
