Rudra Pratap Singh

A support inbox that sorts every ticket by team and urgency. A small model I fine-tuned handles the easy ones in milliseconds; only the tickets it is unsure about go to an LLM, and the ticket shows why.

57%of tickets never reach the LLM80.7%accuracy, vs 81.0% LLM-only8 msper ticket on one CPU core23 MBmodel after int8

How it works

• Distilled, then calibrated: MiniLM fine-tuned on 18,000 real support tweets labelled by an LLM, with two heads (team and urgency); temperature scaling cut calibration error from 0.108 to 0.066, so its confidence can decide when to escalate.
• Fixed a real export bug: the int8 ONNX model matched PyTorch on only 92% of tickets; the old exporter had baked the sequence length into the mask. Moving to the new exporter brought it back in line.
• Built like a service: BullMQ queue on Redis, live updates over WebSockets, a circuit breaker when the LLM quota runs out, and a Next.js dashboard with a results page. Tried a reranker and request batching, measured both, and dropped both.

Python · PyTorch · ONNX Runtime · FastAPI · Gemini API · Node.js · BullMQ · Redis · WebSockets · Next.js · TypeScript · Playwright

Results and design decisions →

A gatekeeper between an AI agent and its tools. It stops poisoned tools, tools that quietly change after approval, and private data leaking to the wrong server, whether the attack is written in English, Hindi, Hinglish or Kannada.

0/25attack runs leaked, with Dwarpal5/25leaked without it0false blocks34tests, CI on Kubernetes

How it works

• One policy graph (LangGraph): tool fingerprints, data-flow tracking across servers, multilingual rules, Prompt Guard 2 and an LLM policy judge; cheap checks first, and a person approves borderline calls.
• Found what classic guards miss: Prompt Guard 2 scored a politely worded attack 0.001 (safe), and the agent followed it and leaked a phone number. Dwarpal stopped it every time.
• Built like a product: Postgres with pgvector, a Redis cache, runs as an MCP server for any client, and a Helm chart tested on Kubernetes in CI.

Python · LangGraph · MCP · FastAPI · PostgreSQL + pgvector · Redis · Docker · Kubernetes · Helm · React

Results and design decisions →

A research assistant for Ayurveda, traditional knowledge and Indian IP law. Ask in English, Hindi, Kannada or Hinglish; it answers only from its sources and cites the exact page behind every claim.

0.99answer faithfulness0.95retrieval Hit@147tests, with CI4languages

How it works

• Retrieval written by hand: section-aware PDF chunks, Gemini embeddings and NumPy cosine search, plus BM25 and RRF built to compare against.
• Decisions from data: hybrid search lowered MRR from 0.97 to 0.93 on 38 labelled questions, so dense search stayed the default.
• Answers judged, and the judge tested: a model from another family checks every claim; it caught all 15 errors I planted.

Python · FastAPI · Gemini API · NumPy · React · Docker · GitHub Actions

Full write-up with architecture →

An AI flat finder for Bengaluru. Describe how you live, and an agent decides which tools to call to find flats that fit, check whether the rent is fair, and compare a shortlist.

20/20tool choice on the eval2.4 smedian replyRefusedprompt injection26unit tests