AI systems built from idea to production deployment.
redcell — Adversarial Test Harness for LLM Agents
Fires real attacks at an LLM app or MCP server and reports what actually got through, organized by the OWASP Top 10 for LLM Applications. Excessive-agency probes are schema-driven: they inspect a target's real tool definitions and verify findings with planted canaries instead of guessing from error text. Found and fixed two real bugs in its own evaluation logic by dogfooding it against a live server.
Security
MCP
Python
pytest
critique-mcp — MCP Server for PR Review
An MCP server that gives an agent the tools to actually review a pull request: lints changed Python files with ruff, flags PRs that change source without touching any test, and posts a real review back to GitHub. Built the exact checks I've done by hand fixing bugs in chroma-core, weaviate, and qdrant's own PRs.
MCP
Python
GitHub API
pytest
nanoglyph — Transformer Built From Scratch
A decoder-only transformer with hand-written causal self-attention: no nn.MultiheadAttention, every matmul and mask is explicit. Trained character-level on Shakespeare's complete works (loss 4.56 → 1.61 over 2000 steps), with a test proving zero future-token leakage rather than just checking output shapes.
PyTorch
Transformers
Streamlit
pytest
HireSmart — AI Recruitment System
End-to-end recruitment platform: job postings, GPT-4 resume screening against the job description, interview panel scheduling and evaluation, personality assessment, and offer letters, with role-based dashboards.
Laravel
GPT-4
Python
FYP
TrainSet AI — Multimodal Dataset Builder
Unified data ingestion pipeline processing audio, images, PDFs, and text to generate structured ML training datasets. Uses LLaMA and LLaVA to convert speech and visual content into clean annotated data.
LLaMA
LLaVA
Multimodal
Python
Semantic Search Engine (LLM-Free)
Fast document retrieval using embedding-based similarity search without large language models. Document chunking and cosine similarity with FAISS for accurate information retrieval via Streamlit.
FAISS
Embeddings
Streamlit
Python
DocIQ — RAG Document Intelligence
Retrieval-augmented document Q&A: FastAPI backend, ChromaDB vector store, Claude for generation. Ships with a retrieval-quality eval harness (hit@3: 90%) that gates CI, and 28 tests exercising real embedding-based retrieval rather than mocks.
RAG
FastAPI
ChromaDB
CI/CD
groundtruth — Research Agent
Plans a question into sub-questions, researches each with live web search, and synthesizes one answer where every claim cites the source that actually supports it, pulled from real search results, never invented.
Agentic AI
Claude API
Python
pytest
agent-pipeline — Multi-Agent Pipeline
Researcher, writer, and critic agents hand off work in sequence with live streaming. The critic checks every claim in the draft actually traces back to the research brief and triggers a real revision when it doesn't.
Next.js
TypeScript
Tailwind
Groq
NPI Physician Data Extraction System
Python-based large-scale data extraction system using the CMS NPI Registry API to collect physician records. Structured validation, retry mechanisms, and request throttling for production reliability.
Python
API
Data Pipeline
Automation