class LanceGalicia(AIEngineer):
 role = "AI Systems Engineer"
 based = "Philippines · GMT+8 · open to remote"
 now = "Building Maxima — a 24/7 production AI companion"
 focus = ["LLM agents", "RAG", "long-horizon memory", "evals", "reliability"]
 stack = ["Python", "FastAPI", "Claude API", "Qdrant", "SQLite", "Three.js"]
 def thesis(self) -> str:
 # The next useful AI layer isn't a bigger chat window.
 return "It's agents that remember, revise, use tools — and can be tested."

📄 Your AI agent's tests are lying to you — five ways I caught mine

I built a checker to catch a corrupted data table. To prove it worked, I corrupted the table on purpose and ran it. It reported clean. It had been reading the wrong path for weeks and swallowing the error — code I wrote myself, with a comment above it saying "never fail the gate on a tooling problem."

Five distinct ways a test can pass while verifying nothing, each a real dated incident from an autonomous loop I run over these repos, with the measurements attached. Ends on the ten-second experiment that finds all five: remove the input and count the findings.

reliability testing ai-agents — Read ›

Each project proves one property production AI actually needs.

🧬 ENGRAM — memory that can be audited A belief-revision ledger + live 3D observatory: what an agent believed, when it changed, what evidence triggered the revision, what decayed, and what was later restored. memory provenance three.js sqlite — Live › · Source

📑 LEMMA — answers that prove themselves Hybrid RAG (dense + BM25 + RRF fusion) with native citations, bounded multi-hop retrieval (search-as-a-tool, ≤3 hops — it decides how hard to look and shows its hops), and a live reliability panel that measures its own recall, latency, and cost. RAG citations multi-hop evals fastapi — Live › · Source

🔁 LOOPKIT — agents that govern themselves A file-based operating system for self-improving autonomous work loops: charter, append-only log, fleet dashboard, golden regression set, and a captions airlock. Its optional stdlib Graph-of-Loops seals immutable reviewed plans, enforces gated edges, waits for named human approval, and resumes without replaying passed side effects. The caller still owns authentication; callbacks are trusted code, not a sandbox. agents autonomy governance stdlib — Source ›

🪚 CARVE — refactors that measure before they move A stdlib-only static planner for splitting large Python modules. It maps coupling, counts dispatch targets separately from physical branches, and sizes proof-gated seams. Unsupported dispatch conditions fail closed, unproven helpers stay welded, and behavioral equivalence still belongs to tests and differentials. static analysis refactoring python ast stdlib — Source ›

⏱️ runproof — silence is a failure mode A job that fails loudly is a good day; the expensive one stops running and says nothing. runproof turns that silence into a verdict — FRESH · LATE · SILENT · NEVER_RAN — from a heartbeat the job writes itself, so "it's been fine" becomes something you check instead of assume. Stdlib-only, no install step. scheduled jobs observability heartbeat stdlib — Source ›

🛡️ Agent Reliability Arena — agents that get tested An eval harness that catches stale memory stated as truth, unsupported tool claims, answer drift, and long-horizon trend health. evals agents observability — Live › · Source

⚙️ FlowProof — automation that fails safely The operational half of reliability: idempotent intake so a retried webhook creates nothing twice, bounded retries that end in an auditable dead-letter state instead of hanging forever, and human approval before ambiguous work runs. Provider-free core — the ledger tests run with no install and no credentials. reliability idempotency audit trail fastapi — Source ›

🤖 Maxima / FORGE — AI that compounds over time Long-horizon companion: persistent memory, current-truth override, tool use, and proactive pattern detection across weeks — not a one-shot chatbot. agents long-horizon memory tools — Live ›

🌌 FORGE Neural Map — systems you can explore A real codebase rendered as a navigable 3D universe — 2,778 nodes · 7,295 connections · 157 systems — vanilla Three.js with GPU-shader layout over a live knowledge graph. three.js dataviz webgl — Live › · Source

🔀 Axiom AI — one door to every model, honest when upstream breaks A unified Claude, GPT, Gemini, and Groq gateway with streaming, auth, usage tracking, and a live Failure Contract Lab: broken streams, missing providers, failed chat turns, and retry ownership made visible from provider-free tests. gateway llm infra fastapi failure contracts — Live › · Source

Python Claude API FastAPI Qdrant SQLite Three.js JavaScript Docker Railway Git Linux

LLM agents · RAG · hybrid retrieval · evaluation harnesses · vector search · observability · tool routing · knowledge graphs · prompt caching · Railway deploys

The next useful AI layer isn't a bigger chat window. It's infrastructure for agents that preserve context, update beliefs when evidence changes, expose why they changed, and never pass stale memory off as current truth. Every project above is a proof of one piece of it.