build log

An incident-response runtime that triages host evidence on the analyst's machine using a local LLM (ONNX, GGUF, or SafeTensors). An agentic ReAct loop iteratively picks investigation tools (logs, listeners, persistence, accounts, processes) and synthesizes severity-scored findings with full audit trails. Supports multiple backends (CPU, CUDA, DirectML, CoreML, TensorRT, QNN, Vitis) and falls back to a deterministic dry-run so triage never stalls when inference fails.

• ▸Bring your own model. No cloud APIs, no data exfiltration, no vendor lock-in.
• ▸Evidence bundles with SHA-256 chain-of-custody and a JSON adapter for SOAR / CI gating.
• ▸Deterministic fallback when live inference fails, with machine-readable reason codes.

A fully offline system to preserve endangered indigenous knowledge in communities where cloud AI isn't an option. A single Gemma model runs voice-guided interviews with elders, identifies plants, tools, and artifacts from photos, and uses native function calling to turn transcripts into structured knowledge-graph entries. Retrieval blends FTS5 keyword search with ChromaDB semantic search, and every answer cites the elder it came from and surfaces safety warnings. Ships an Unsloth LoRA fine-tuning recipe and a Gemma Edge fallback for mobile-class hardware.

• ▸100% offline inference. No cloud APIs or keys; one Gemma model handles reasoning, vision, and extraction.
• ▸Hybrid retrieval (FTS5 keyword + ChromaDB semantic) with per-elder attribution and safety surfacing.
• ▸Offline speech-to-text (Whisper) and text-to-speech in 100+ languages (espeak-ng).

A Python wrapper around the C reference implementation of FAEST, a NIST post-quantum signature candidate built on symmetric primitives (AES-VOLE) instead of structured lattices. Ships all 12 FAEST parameter sets with automatic memory-clearing of private keys. Prebuilt wheels for Linux x86_64/aarch64 and macOS arm64/x86_64 across Python 3.8 to 3.13, so consumers don't need a C toolchain.

• ▸All 12 FAEST parameter sets (128/192/256-bit, fast vs small variants, EM mode).
• ▸Prebuilt wheels. No compilation step for end users.
• ▸Pythonic API with typed exceptions for keygen / sign / verify failure paths.

A disaster-relief intake platform built for IBM's AI Experiential Learning Lab that meets citizens wherever their connectivity allows. Three surfaces share one agent workflow: a universal web intake form, an SMS path for low-bandwidth and feature-phone users, and a caseworker triage dashboard. An agentWorkflow service layer wraps IBM watsonx.ai Granite models and Orchestrate skills behind a deterministic mock mode, so demos stay reliable with zero credentials while real models drop in via env vars.

• ▸Three intake channels (web, SMS, agency dashboard) over a single agent workflow.
• ▸watsonx Granite + Orchestrate hooks with a credential-free mock mode for reliable demos.
• ▸Netlify serverless deployment with a live public POC.

Led a 4-person team to build an end-to-end AI meeting-minutes generator. Fine-tuned BART-large on AMI and ICSI datasets and extended the effective context from 1024 to 8000+ tokens using n-gram overlap stitching. Deployed as microservices behind a Node.js/Puppeteer bot that joins Google Meet calls and streams transcripts in real time.

• ▸Context extension from 1024 to 8000+ tokens via n-gram overlap.
• ▸Autonomous Google Meet bot (Node + Puppeteer).
• ▸2 copyrights + merit position at Aavishkar Research Convention 2024.

Smart India Hackathon prototype. Built a system for law enforcement to trace the origin and propagation of flagged electronic messages using ECDSA digital signatures anchored to a blockchain. Wrote the API gateway for smart-contract integration and a React dashboard for forensics analysts.

• ▸ECDSA-signed envelopes anchored on-chain.
• ▸API gateway bridging Node and Solidity contracts.
• ▸Poster at Conference on Technologies for Future Cities 2022.