Projects

Selected work — MCP security, production voice AI, and low-resource-language Document AI.

Projects

A few things I’ve built. Client engagements are described without identifying details.

Provenire — MCP security → compliance evidence

An open-core scanner for the Model Context Protocol attack surface. It connects to an MCP server, detects security weaknesses in the tools it exposes (prompt-injection directives, invisible-character smuggling, exfiltration affordances, over-privilege, unbounded schemas), scores them, and — the part nobody else does — maps those findings to named compliance controls and emits a deterministic, signed evidence record. Built strictly test-first; the engine and CLI are Apache-2.0.

Architecture: framework-neutral detection engine + regulation packs as data + evidence bundle. “Detection ≠ mapping.”

Read the write-up · GitHub

Low-latency real-time voice AI agent

A production, real-time voice AI agent built on a six-layer LLM-native stack (telephony → speech-to-text → LLM → text-to-speech → guardrails), tuned to sub-second response latency with safety enforced by policy guardrails. Covered regression suites, latency reviews, and a full agentic-workflow design. (Client details omitted.)

Stack: LiveKit · Deepgram Nova-3 · GPT-class LLM · ElevenLabs · NeMo Guardrails.

Odia OCR & low-resource-language Document AI

A document-understanding pipeline for Odia, a language most OCR and OCR-LLM systems handle poorly — chaining modern OCR and vision-language models over curated Odia datasets, with an interactive breakdown of the model internals (a from-scratch decoder and a tokenizer built for Odia conjunct consonants). A portfolio project about making capable AI work for languages the mainstream ignores.

Stack: DeepSeek-OCR → olmOCR → Qwen2.5-VL; RunPod serverless; Hugging Face Spaces demo.


More on the blog, and code on GitHub.