A 2026 agentic, retrieval-grounded AI career assistant for HBCU students — model-agnostic (Claude by default), with hybrid RAG and a Model Context Protocol server. Evolved from IgniteAI, the ChatGPT custom GPT my team built and took to 🥇 first place at the HP Future of Work Accelerator 2024 (HBCU Technology Conference).
▶ Live demo → lodestar.sanyer.org — chat with the agent live (Claude Haiku, rate-limited). The hosted build is a lean Flask + BM25 deployment; the full FastAPI + hybrid-RAG (LanceDB) stack below runs locally and in Docker.
At a glance: model-agnostic LLM layer (Claude · OpenAI) · hybrid RAG (BM25 + LanceDB vectors, fused with RRF) · agentic tool-use (router → tool loop) · exposed as an MCP server for Claude Desktop · LLM-as-judge evals · OWASP/FERPA guardrails · FastAPI web UI (deploy-ready) · 36 tests · ruff-clean · Python 3.14 / uv · offline test mode
In 2024, my HP FOWA team built IgniteAI — a ChatGPT custom GPT that coaches HBCU students on résumés, interviews, internships, and scholarships — and won first place. In 2026 I rebuilt it as Lodestar: a real, runnable agentic RAG system that decides when to retrieve, grounds its answers in a knowledge base, runs model-agnostic (default Claude), and plugs into Claude Desktop as an MCP server. This repo ships both — the modern rebuild and a faithful reproduction of the original — side by side, so the 2024 → 2026 evolution is visible in one place.
flowchart TD
U([HBCU student]) --> ENTRY["lodestar CLI · MCP client (Claude Desktop)"]
ENTRY --> ROUTER{"Router — classify intent<br/>(résumé · interview · scholarship · ...)"}
ROUTER --> AGENT["Agent loop — Claude native tool-use"]
AGENT -->|"tool: retrieve_knowledge"| HYB["Hybrid retrieval<br/>BM25 (sparse) + LanceDB·fastembed (dense) → RRF fusion"]
AGENT -->|"tool: web_search"| WEB["Web search"]
HYB -->|grounded context| AGENT
WEB -->|results| AGENT
AGENT --> ANS(["Grounded, actionable answer"])
AGENT -.->|model-agnostic provider| LLM[("Claude · OpenAI")]
The same engine is exposed two ways: a CLI (uv run lodestar) and an MCP server
(uv run python -m lodestar.mcp_server) that registers retrieve_knowledge + get_career_advice
tools into Claude Desktop. With TEST_MODE=true everything runs offline (mock model + deterministic
embeddings) — no key required.
Transcript — “How do I find scholarships for HBCU STEM students?”
$ uv run lodestar # TEST_MODE=false, ANTHROPIC_API_KEY set
Lodestar (agentic) — live (claude-sonnet-4-6). Type 'quit' to exit.
You: How do I find scholarships for HBCU STEM students? Give 2-3 concrete next steps.
Lodestar:
## 🎓 Finding Scholarships as an HBCU STEM Student
### Where to Look — real programs to target
| Organization | What they offer |
|---|---|
| UNCF (United Negro College Fund) | STEM-focused scholarships — uncf.org/scholarships |
| Thurgood Marshall College Fund | Awards for HBCU students, incl. STEM — tmcf.org |
| NSBE, SMART (DoD), NSF S-STEM | Society + federal STEM funding pipelines |
### Your next steps
1. Build a one-page tracker (deadline · GPA · essays · recommenders).
2. Apply to 2–3 per week; ask your department chair about named/funded research.
3. Start with UNCF + TMCF this week — they cover the widest set of HBCU STEM awards.
(The agent routed the question, called retrieve_knowledge, grounded on the knowledge base,
and synthesized this with claude-sonnet-4-6. Generate your own recording — see
docs/media-guide.md.)
| Capability | How it's built |
|---|---|
| Model-agnostic LLM | LLMProvider interface + AnthropicProvider (default, Claude, prompt-caching), optional OpenAIProvider, offline MockProvider |
| Hybrid RAG | fastembed dense embeddings in a LanceDB vector store + BM25 sparse, fused via Reciprocal Rank Fusion (retrieval/) |
| Agentic tool-use | keyword router → Claude native tool-use loop with retrieve_knowledge + web_search (agents/) |
| MCP server | FastMCP server exposing Lodestar's tools to Claude Desktop / any MCP client (mcp_server.py) |
| Evals | LLM-as-judge harness scoring relevance/accuracy/actionability/safety (evals/) |
| Security | OWASP-LLM-Top-10 mapping, prompt-injection guardrail, PII detection (safety.py, docs/security.md) |
| Resilience | graceful BM25 fallback if the embedding model can't be fetched; full offline test mode |
Latest LLM-as-judge eval (Claude-judged, 4 representative queries): relevance, accuracy, and actionability 5/5; safety 4–5/5. Raw runs in
evals/runs/.
| Package | What it is | Run it | |
|---|---|---|---|
| Lodestar (2026) | src/lodestar/ |
the modern agentic-RAG + MCP rebuild | uv run lodestar |
| IgniteAI (2024) | src/ignite/ |
faithful reproduction of the first-place custom GPT (OpenAI gpt-4o-mini) |
uv run ignite |
Both are clean-room — written fresh from the team's architecture docs + the IgniteAI
specification, not the original GPL-3.0 team code (see docs/decisions.md).
The original product spec is preserved in product/.
uv sync --extra dev # create env + install deps (Python 3.14)
cp .env.example .env # TEST_MODE=true runs offline; add ANTHROPIC_API_KEY for live Claude
uv run lodestar # the agentic Lodestar assistant
uv run ignite # the 2024 IgniteAI reproduction
uv run pytest -m unit # 33 offline tests
uv run python -m lodestar.mcp_server # serve over MCP (see docs/mcp.md to wire into Claude Desktop)A FastAPI single-page chat ships in the box (src/lodestar/web.py + static/):
uv run lodestar-web # http://127.0.0.1:8000 (offline mock by default)Containerized (Dockerfile) and one-command-deployable to Railway / Fly /
Render — set ANTHROPIC_API_KEY + TEST_MODE=false for live Claude. See
docs/deploy.md.
Scope was driven by IBM Technology's "AI Periodic Table" curriculum — every track maps to
something concrete here: RAG → hybrid retrieval, Agents → the router + tool loop, MCP → the server,
AI Security → the guardrails, Governance → the evals. Full mapping:
docs/ibm-curriculum-mapping.md.
Python 3.14 · uv (env + lock) · Anthropic SDK (Claude, default) · fastembed (ONNX embeddings) · LanceDB (vector store) · rank-bm25 (sparse) · MCP Python SDK · ruff · pytest · cryptography (Fernet, FERPA demo).
lodestar/
├── src/lodestar/ # 2026 rebuild — providers/ retrieval/ agents/ prompts/ mcp_server.py · CLI: lodestar
├── src/ignite/ # 2024 IgniteAI reproduction · CLI: ignite
├── product/ # IgniteAI custom-GPT spec + knowledge (what shipped in 2024)
├── data/ # sample knowledge base (offline demo/tests)
├── evals/ # LLM-as-judge evaluation harness
├── docs/ # architecture · decisions · roadmap · mcp · security · IBM-curriculum · media-guide
├── competitions/ # HP FOWA 2024 first-place record
├── portfolio/ # case study (problem → solution → contributions → result)
├── assets/ # README media (infographic / video / demo GIF)
└── tests/ # unit (offline) + integration (live, auto-skip)
| Doc | What |
|---|---|
docs/roadmap.md |
The two-track design |
docs/architecture.md |
System architecture (the production vision + what's built) |
docs/ibm-curriculum-mapping.md |
IBM "AI Periodic Table" → code |
docs/mcp.md |
Wire the MCP server into Claude Desktop |
docs/security.md |
OWASP LLM Top-10 + FERPA posture |
docs/decisions.md |
Clean-room boundary + design decisions |
portfolio/case-study.md |
The HP FOWA story + my contributions |
IgniteAI was a team effort by HP FOWA 2024 Team 4 — see NOTICE.md for attribution. This repository (the Lodestar rebuild + the reproduction + all docs) is authored by Wolfgang Sanyer and licensed under the Polyform Noncommercial License 1.0.0 (LICENSE) — free to view, study, and share for non-commercial purposes.
Wolfgang Sanyer · wolfgang.sanyer@gmail.com