// before the first model call
Six years of distributed systems, at scale.
Banks, federal auditors, global manufacturing, energy. Java and distributed systems long before AI, which is exactly why the AI I ship has to survive production.
2025-26Avangrid · Senior AI Engineer. RAG diagnostics agent over 400+ microservices; mean-time-to-investigate down 80%, incident detection 60min to under 10.
2025Yubico · Senior Software Engineer. Led PingOne OAuth2/OIDC integration end to end; automated test coverage 40% to 90%.
2025Estée Lauder · Senior Software Engineer. Cut API response times 70% on a platform serving 1M+ requests/day.
2023-25Tribunal de Contas da União · Senior Full Stack. Procurement transparency for 30,000+ federal auditors; data access 2min to 10s; 20+ SQL sources into one index of 10M+ records.
2022-23CNH · Full Stack. Authz/RBAC for 60,000+ employees across 180 countries; critical defect rate down 30%.
2020-22Banco do Brasil · Full Stack. Quarkus microservices for legislative records; operational efficiency up 60%, CI/CD success rate up 40%.
// the flagship
Shared memory for AI coding agents
Switch from Claude Code to Codex to Cursor, or come back to a project after a week,
and your assistant is blank. Every decision, every dead end, gone. So you re-explain.
Every session.
AXON catches context at the moment it forms, on git commits and session
boundaries, and hands it back on demand over MCP, with a plain .axon/context.md
fallback. Same memory, any agent. You resume instead of restart.
Fix the cause, not the symptom
A backend engineer would never fix a 500 by rewriting the error message. But
re-explaining context to an agent every morning is that exact move: patch the
symptom, skip the cause. AXON is the cause fixed once.
Underneath: a graph plus a vector store on PostgreSQL and pgvector, graph retrieval
delegated to GLYPH, and every decision written as an ADR instead of
buried in the diff. Destructive tools ask before they run.
The scar
I shipped a “52% token savings” number. It read like a result. It was a
deterministic model, and I let it sit on my README and my LinkedIn for months.
So I built GNOMON and ran the real A/B, recall on versus off.
The honest answer is not the flattering one: recall lifts answer faithfulness
from about 0.46 to about 0.74, and it costs roughly 2,151 extra tokens a
turn. It buys quality, not a discount. I retired the 52%.
The savings claim didn’t disappear, it got replaced with one I can point at: in
production, across 100 real requests, AXON reads 90.7% less than the
alternative, about 32,000 tokens instead of 346,000, measured against actually
reading each source file in full. Different question from the recall number
above, same rule: reproduce it or don’t say it.
That is the whole point of this page. A number you can reproduce beats a number
that sounds good.
// the real, measured results
Two numbers, both measured.
AXON recall / measured A/BGNOMON · 2026-07-02
0.46 → 0.74answer faithfulness, recall off vs on. 95% CI, intervals do not overlap.
measured, provider-meteredN = 17 casescost: +2,151 input tokens / turn
// a single-turn quality lift and its real token cost, not a savings claim.
AXON production reads / measuredaxon scripts/recall_savings_report.py · 2026-07-04
90.7%fewer tokens read vs. loading each source file in full, real production requests.
measured, per-chunk telemetryN = 100 requests32,298 vs 346,081 tokens
// a read-volume result, different question from the faithfulness card; don't blend them.
// the system AXON anchors
AXON does not work alone.
It sits at the centre. Three feed it or measure it directly; the fourth is what one of those tools was built to prove out. All real, all smaller than AXON on purpose, so the flagship stays the flagship.
GLYPH
Graph-aware retrieval. It finds context by following the relationships between code and docs, where plain similarity search goes blind. Its community summaries answer "how is this organized?" at half the tokens, same quality.
I built it expecting the graph to beat plain vector search. On code, it lost. I published the negative result instead of hiding it.
go deeper →GNOMON
Offline RAG eval harness. Every score comes with a confidence interval, cost, and latency.
It measured AXON's recall (faithfulness 0.46 to 0.74) and it is what retired my modeled 52%.
go deeper →rpg-master-ai
Production RAG in Java 21 and Spring AI. Ingests RPG rulebooks and answers rules questions in English and Portuguese, with page-level source attribution.
GNOMON caught the eval lying here first: a deterministic judge scored N times looked like N independent cases to the naive math. It was not. Fixed the confidence interval before trusting the score.
go deeper →rtkx
A token-optimising CLI proxy. It compresses repetitive dev-command output before it reaches the model.
The compression layer AXON leans on for how compact, while GLYPH answers what context.
go deeper →