Anant Patankar · India · Open to ML/AI roles
ML/AI Engineer — Agentic Systems & Generative AI
Multi-agent orchestration (Google ADK) · MCP Protocol · Production RAG · MLOps
Forensic scientist turned ML engineer — I bring evidence-grade rigor to agentic AI and grounded retrieval.
What I do
Agentic AI
Multi-agent orchestration with Google ADK, MCP protocol servers, and LLM-driven reasoning & tool use — systems where the model decides, and the architecture keeps it safe, cheap, and auditable.
Production RAG
Grounded generation with source traceability, dense and hybrid retrieval, and cross-encoder reranking — measured with real recall and latency numbers, not vibes.
Production Systems & MLOps
Async Python, distributed pipelines (Celery/Redis/Cloud Run), containerized deploys, experiment tracking and model registries — the backbone that makes the AI shippable.
Skills
Tech Stack
01 / Agentic AI & LLMs
02 / Vector Search & Retrieval
03 / ML / Deep Learning
04 / Distributed Systems & Infra
05 / MLOps & Cloud
06 / Data Engineering
Selected work
Featured Projects
Five systems spanning agentic engines, MCP servers, and RAG pipelines. Two are open source and independently verifiable.
FLAME — Agentic Recommendation Engine
Google ADK · Gemini 2.5 Flash · GCP · FlagshipAI backend for a gamified crowd-labelling platform that turns plant workers into a distributed workforce annotating workplace-safety imagery through four mini-games. A Google ADK agent drives LLM-based topic selection and search-query authoring; deterministic Python owns bucket rotation, consensus, Elo ranking, trust scoring, and anti-gaming.
12-agent engine · 5 ADK skills · designed for 100K+ concurrent users & 50K+ images/day
AI-Powered HSE Training Content Generation
Vertex AI Agent Engine · RAG · Human-in-the-loop14-agent content-generation pipeline that turns safety documents into reviewed microlearning — grounded RAG with per-question source attribution and a structural guarantee that nothing publishes without human approval.
14 agents · multi-format RAG ingestion · zero auto-publish path
Lumenore Analytics MCP Server
MCP Protocol · Async PythonProduction MCP server exposing a BI platform’s analytics to any AI assistant — natural-language querying, forecasting, correlations, outlier detection — over typed tools with per-request credential isolation and no server-side state.
7 typed analytics tools · connection-pooled async client · MIT
MLOps MCP Server
MCP Protocol · MLOps · Personal OSSOpen-source MCP server giving AI assistants tool-based access to MLOps workflows — experiment tracking, model registry, dataset drift detection, pipeline DAGs, lineage tracing — wrapping MLflow, DVC, and Git.
pip install mlops-mcp-server · v0.1.0 · MIT · two-tier tool registry
Enterprise Knowledge Q&A
Production RAG · Retrieval · EvaluationProduction RAG over enterprise document libraries: ChromaDB dense retrieval with metadata-filtered per-business-unit isolation and token-budgeted, cited generation — turning ~15-minute manual document hunts into ~30-second answers.
85%+ Recall@5 (200 labelled queries) · ~20% token-cost reduction · 200+ users across 3 BUs
Verifiable proof
Open Source
Don’t take my word for it — read the code.
mlops-mcp-server
PyPIMCP server for MLOps workflows: experiment tracking, model registry, KS-test drift detection, DAG cycle detection (Kahn’s algorithm), BFS lineage tracing with Mermaid visualizations.
lumenore-mcp
Org repoProduction analytics MCP server I authored at Netlink — FastMCP, async aiohttp with connection pooling, per-request credential pass-through, no server-side state. Published open source under the Lumenore Platform org.
Track record
Experience
Architected the agentic AI backend (FLAME) for an enterprise HSE gamification platform and designed a 14-agent training-content generation pipeline — Google ADK, Vertex AI, Gemini, Cloud Run. Led technical direction for a junior engineer; presented vector-DB trade-off analysis to senior stakeholders.
Built the open-source Lumenore Analytics MCP Server; an AI customer-support chatbot with Qdrant hybrid search + cross-encoder reranking (85%+ precision@5, sub-200ms P95 accept/ack at 10K+ daily messages); and an agentic MCP client/orchestrator managing multiple MCP servers across 4 transport protocols with RBAC/ABAC (~55% latency reduction on high-fan-out queries).
Built the retrieval components of a production Enterprise Knowledge Q&A RAG system (85%+ Recall@5); a code-analysis AI processing 500K+ lines across 10+ languages (drove a 65% code-review-time reduction); and ~10–12 of a 30+ use-case ML/DL portfolio for a rural-development bank on 10M+ records. 2× Game Changer of the Month (Sept & Dec 2023).
Built RF/SVM models for reactor-parameter optimization and an LSTM-autoencoder for real-time anomaly detection, cutting material waste ~12%.
Built an ML forecasting pipeline (ARIMA, RF, LSTM) and a real-time object-detection system processing 15 FPS video with TensorFlow.
About
From forensic science to agentic AI
I hold an M.Sc. in Forensic Science and moved into machine learning through self-directed study, building toward production ML and AI over nearly six years — the last ~3.5 in Generative AI. The analytical discipline transferred directly: I design agentic and RAG systems with an emphasis on auditability, grounding, and measurable outcomes.
My current focus is multi-agent orchestration (Google ADK), MCP-based tooling, and production RAG, supported by distributed-systems and MLOps practices — with two open-source MCP servers published publicly.
Get in touch
Contact
Currently open to ML/AI Engineer roles and available to join immediately — happy to talk agentic AI, GenAI systems, and production ML.