I build production AI systems from scratch — not wrappers around APIs, not demos, not prototypes.

Full autonomous runtimes: multi-agent orchestration, FSM lifecycle enforcement, closed-loop self-healing observability, semantic memory, evidence-based verification, real-time dashboards, and token-cost-aware model routing. Every system below is live.

15 years in software engineering. AI-native infrastructure since 2022. Currently building VRASHOWS — a strategic AI partnership for companies that want serious AI, not slide decks.

VRAXIA is a cognitive runtime I designed, built, and operate as a solo engineer. It is not a product demo. It is a production multi-agent OS that I use daily and continuously evolve. Every component is documented, tested, and architected to be portable across industry verticals.

Three public systems. One shared cognitive core.

github.com/SAMIRRICARDO/vraxia-work-system

"Not a job bot. A multi-agent cognitive runtime that happens to apply to jobs."

The most technically comprehensive component of VRAXIA. 20+ specialized agents, a 5-layer answer resolution pipeline, real-time observability with AI self-healing, FSM lifecycle enforcement, and a live operations dashboard — all built solo in TypeScript.

┌──────────────────────────────────────────────────────────────────────┐
│ VRAXIA WORK RUNTIME │
├─────────────────────────┬────────────────────────┬───────────────────┤
│ PERCEPTION LAYER │ INTELLIGENCE LAYER │ MEMORY LAYER │
│ │ │ │
│ OtelAgent │ 20+ AI Agents │ CKOS (RAG) │
│ ├─ Metric ring buf. │ ├─ Questionnaire │ 754 chunks │
│ ├─ SSE fan-out │ ├─ ProfessionalTwins │ 41 md sources │
│ ├─ Span/log ingest │ ├─ HireScoreEngine │ TF-IDF + BM25 │
│ └─ Anomaly detection │ ├─ ATSOptimizer │ │
│ │ ├─ ResumeAgent │ Career Memory │
│ OtelSentinel │ ├─ MatchAgent │ Redis episodic │
│ ├─ reading_insight │ ├─ SalaryAdvisor │ pgvector LTM │
│ ├─ picking_action │ ├─ InterviewCoach │ │
│ ├─ executing │ ├─ NetworkingAgent │ CandidateKB │
│ └─ resolved/failed │ ├─ ProfileGapAnalyzer │ Skill normaliz. │
│ │ ├─ LearningAgent │ Gap detection │
│ SenseBridge │ ├─ TwinSelector │ │
│ Incident lifecycle │ ├─ RequirementsAgent │ │
│ │ ├─ ProfileOptimizer │ │
├─────────────────────────┤ ├─ ProfileAutoUpdater │ │
│ EXECUTION LAYER │ ├─ StatusTracker │ │
│ │ └─ ATSAgent ├───────────────────┤
│ ApplicationStateMach. │ │ MARKETPLACE │
│ 18 lifecycle states │ ChatOrchestrator │ │
│ 17 error categories │ Intent classification │ Plugin API │
│ │ 9 intent domains │ CoverLetter │
│ ApplicationTruthEng. │ │ LinkedInOptim. │
│ Multi-source evidence ├────────────────────────┤ HeadhunterScript│
│ Confidence scoring │ ORCHESTRATION LAYER │ StartupRadar │
│ │ │ EquityCalc. │
│ EvidenceCollector │ Scheduler + DAG │ VisaFilter │
│ RetryEngine (17 cat.) │ Daily automation │ │
│ ValidationEngine │ Telegram notify │ BYOK ready │
│ ErrorClassifier │ Remote dev agent │ │
└─────────────────────────┴────────────────────────┴───────────────────┘

The most differentiated subsystem: a self-healing observability loop where anomalies trigger LLM analysis (Haiku, <800 tokens), map to discrete remediation actions, execute autonomously, and broadcast every step live to the dashboard via SSE.

Anomaly detected (error rate, latency, queue depth)
│
├─► SenseBridge opens incident
│
└─► OtelSentinel.triage()
│
├─ [SSE → dashboard: reading_insight]
│ OtelAnalyst (Claude Haiku) analyzes snapshot + history
│ Returns: severity, summary, interventions[], tokensUsed
│
├─ [SSE → dashboard: picking_action]
│ intervention keywords → RemediationKind mapping
│ restart_server | drain_cache | wait_and_recheck | alert_only
│
├─ [SSE → dashboard: executing]
│ Action runs: port kill → process spawn → health check
│ OR: 30s wait → recheck metrics
│
└─ [SSE → dashboard: resolved / failed]
Incident closed + Telegram notification
TruthRecord written for audit

No human in the loop. No polling. Dashboard robot panel updates phase by phase in real time.

When the system encounters a form question it has never seen, it does not call the LLM immediately. It works through a cost-optimized resolution stack:

Layer 1 → Cache hit (0 tokens, <1ms)
Layer 2 → Semantic retrieval (CKOS: 754 chunks, TF-IDF scoring)
Layer 3 → Database lookup (structured candidate profile)
Layer 4 → LLM synthesis (Haiku, <400 tokens, JSON output)
Layer 5 → Reasoning fallback (Sonnet, only if Haiku fails)
Result: >85% of questions resolved without LLM call

18 application states. 17 error categories. Zero illegal transitions at runtime.

discovered → queued → starting → opening_job → opening_easy_apply
→ uploading_resume → filling_questions → reviewing → submitting
→ submitted → validating → confirmed ✓
↓
rejected / interview / offer / hired

Error categories: DOM_ERROR · LOGIN_ERROR · CAPTCHA_ERROR · SESSION_ERROR · TIMEOUT_ERROR · ATS_ERROR · UPLOAD_ERROR · SUBMIT_ERROR · LLM_ERROR · NAVIGATION_ERROR · OAUTH_ERROR · TOKEN_ERROR · DATABASE_ERROR · API_ERROR · RATE_LIMIT_ERROR · ANTI_BOT_ERROR

Each category carries an isRetryable flag and configurable retry policy.

Task type → Model → Cost/1k tokens
─────────────────────────────────────────────────
Filter / classify → Haiku → $0.00025
Questionnaire fill → Haiku → $0.00025
Observability LLM → Haiku → $0.00025
Orchestration → Sonnet → $0.003
Planning/reflection→ Opus → $0.015
Average: ~$0.000034/event across full pipeline

The intelligence layer powering the VRAXIA WORK dashboard. A dedicated Report Intelligence Framework (RIF) runs 13 specialized AI agents sequentially, synthesizes their outputs into a structured ComposedReport, and serves a flat intelligence map to the Career OS panel in real time.

Weekly RIF Pipeline — 13 agents, 1 composed report
────────────────────────────────────────────────────────────────────────
executive-report-agent → exec summary + KPIs
pipeline-audit-agent → application pipeline state
winning-stack-agent → tech stack match analysis
winning-profile-agent → careerDNA · roleAffinity · companyAffinity
skills-gap-agent → topSkills [{skill, level, demand}]
root-cause-agent → whySummary (WHY not getting interviews)
topROIActions [{action, expectedGain, effort}]
company-intelligence-agent → company aderência scoring
interview-intelligence-agent → interview pattern analysis
market-intelligence-agent → market signal synthesis
career-audit-agent → career phase + trajectory
learning-engine-audit-agent → learned patterns from historical data
forecast-agent → next-week projections
recommendation-agent → prioritized action list

Every dashboard card implements a mandatory fallback chain — no card ever shows a bare —:

1. Primary value (real data from composed report)
2. Derived fallback (alternative field from same report section)
3. Generic fallback (cross-section derived value)
4. Explanatory message (why data is missing + what to do)

This means the dashboard is always informative — even before the first report is generated.

ialeadsnow.vercel.app · github.com/SAMIRRICARDO/vraxia-system-os

"AI-driven lead generation with real-time pipeline monitoring and autonomous outbound enrichment."

IALEADSNOW is the commercial application layer of VRAXIA — a live B2B lead intelligence platform with real-time AI agent monitoring, multi-department KPIs, and outbound operations at scale.

Next.js 15 (App Router) · TypeScript strict · Tailwind CSS v4
shadcn/ui + Radix primitives · Recharts 3 · SSE real-time streams
Vercel Edge · per-domain layout isolation · glass-morphism design system

github.com/SAMIRRICARDO/vraxia-sense

"Traditional AI responds when you ask. VRAXIA Sense perceives before you notice."

A three-tier event processing pipeline that sits between the external world and your AI agents. It classifies signals autonomously, eliminates noise at zero cost, and notifies humans only when action is required.

LEVEL 0 — Deterministic filter $0.0000/event → 76% of events eliminated
LEVEL 1 — Haiku triage ~$0.00008/evt → binary: relevant?
LEVEL 2 — Full LLM classification ~$0.00030/evt → intent + action + handoff
NOTIFICATION — Telegram only on handoff=true

First pilot: LinkedIn reply classification for B2B outbound (Waalaxy integration).

Every VRAXIA component is domain-agnostic by design. The abstractions are industry-neutral; only the agents' prompts and knowledge base change.

The VRAXIA OS architecture covers all major enterprise departments with dedicated agent logic. Each department is a pluggable module with its own agent, prompt, and data contract:

Sales & Commercial — Lead scoring · pipeline automation · deal stage prediction · next-action engine · revenue forecasting

HR & Talent — Candidate screening · interview coaching · compensation benchmarking · skills gap analysis · headcount planning

Marketing — Email campaign intelligence · audience segmentation · funnel analysis · content optimization · outbound personalization

Finance — Cash flow monitoring · cost-per-agent tracking · burn rate analysis · revenue trend detection · budget anomalies

Operations — System observability · incident lifecycle · self-healing automation · SLA monitoring · process orchestration

Legal — Document review prep · requirements extraction · compliance flagging · contract structuring assistance

Customer Success — Intent classification · churn signal detection · support ticket triage · escalation routing

Executive — Cross-department KPI aggregation · strategic signal monitoring · decision-support synthesis

Traditional RAG retrieves what was written. Human RAG maps how a person thinks, decides, and prioritizes — encoding reasoning patterns, decision trees, and tacit knowledge into queryable structure that an AI agent can navigate.

Implemented in VRAXIA OS: 8 departmental agents, 1,161 indexed skills. Implemented in CKOS: 754 chunks from 41 sources, TF-IDF retrieval with BM25-style scoring, role-aware context injection.

Publication:O Maior Ativo da Sua Empresa — Amazon KDP · June 2026. First Brazilian book on Human RAG applied to organizational AI and knowledge preservation.

Patterns & Principles

Finite State Machines · Multi-agent orchestration · Closed-loop self-healing · Local RAG (TF-IDF + BM25) · SSE real-time streaming · BYOK multi-tenant · Token-cost routing · Evidence-based verification · FSM lifecycle enforcement · Intent classification · Plugin marketplace · Human RAG · Episodic memory · Semantic vector search · Proactive perception

If your company needs any of the following — reach out:

• Autonomous agent systems with full lifecycle enforcement, audit trails, and evidence capture
• Multi-agent orchestration with cost-optimized model routing (Haiku → Sonnet → Opus)
• Production observability with AI-powered anomaly detection and self-healing loops
• RAG pipelines that encode human reasoning, not just documents
• Commercial AI platforms with lead intelligence, pipeline automation, and real-time dashboards
• AI-first architecture for any industry vertical — the domain changes, the architecture holds