CHINMAY HEBBAL.

Built an MCP server that exposes Gmail as a set of AI-native tools. Any MCP-compatible agent — including Anthropic's Claude — can read, search, filter and send emails through standardised tool calls over IMAP/SMTP, without any browser automation. Ships with a FastAPI-backed web UI for direct inbox management and a full Docker setup for one-command deployment.

• 8 MCP tools: list, search by sender/subject, get unread, get details, send, list folders
• Two deployment modes: browser-based web UI (port 8000) or headless MCP server
• App Password auth — no OAuth flow, no browser dependency
• Fully Dockerised with compose file for both modes

A multi-agent pipeline where four specialised agents coordinate over Google's A2A protocol to generate structured courses on any topic. The Orchestrator routes tasks to a Researcher (Google Search), a Judge (pass/fail quality gate), and a Content Builder — iterating up to three times until the research meets the bar, then streaming the final course directly to the browser via SSE.

• 4 agents: Researcher, Judge, Content Builder, Orchestrator — each a separate microservice
• Quality loop: Researcher retries up to 3× based on Judge feedback before handoff
• Real-time SSE streaming from Orchestrator to frontend as course is written
• All agents powered by Gemini 2.5 Pro with Google Search grounding

A rigorous benchmarking framework for small language models running locally via Ollama, targeting a 4 GB VRAM machine (RTX 3050 Laptop). Measures time-to-first-token, throughput (tokens/sec), and quality across 6 prompt categories using 5 distinct scoring strategies. Results surface in a live Streamlit dashboard with radar charts, heatmaps, and side-by-side model comparison. Qwen2.5:3B ranks #1 on quality; Gemma2:2B leads on speed.

• 3 models: Qwen2.5:3B, Gemma2:2B, Llama3.2 — all run on 4 GB VRAM
• 20 prompts × 6 categories: reasoning, coding, factual, maths, summarisation, instruction-following
• 5 scorers: keyword overlap, exact match, code fence, structural (word/line count), ROUGE-1
• Dashboard: radar chart, quality heatmap, TTFT box plot, history browser

End-to-end benchmark of the three leading LLM serving backends across practical, overload, and extreme concurrency levels. Measures RPS, TPS, TTFT, and inter-token latency alongside quality evaluation (MMLU, GSM8K, HumanEval) with a Streamlit dashboard and Parquet export. TRT-LLM delivers 41% more TPS than vLLM at saturation; vLLM and SGLang reclaim the lead when the queue never empties.

• 3 backends: vLLM (FlashInfer), SGLang, TRT-LLM (compiled TRT engine, CUDA 12.9)
• 4 test suites: baseline sweep, fine-grained saturation, overload, and extreme overload up to c=5,000
• Quality eval: MMLU, GSM8K, HumanEval — direct-API fallback when GuideLLM is absent
• Streamlit dashboard: KPI cards, TPS/TTFT curves, tradeoff scatter, raw data export