> ## Documentation Index
> Fetch the complete documentation index at: https://aiescu.com/llms.txt
> Use this file to discover all available pages before exploring further.

# The operator brief: how AIESCU ships production AI

> AIESCU's operator brief, by Teo Deleanu, AIESCU LLC. v2.0, 2026-09-25. Public version: 4 private sections (architecture, ADRs, and runbook excerpts) open only with a private access link and are not included here.

Book a 30-min scoping call: https://cal.com/aiescu (direct with Teo: scoping, references, or questions about a sprint).

## 00 · Overview

**Production-grade RAG for PDFs, scans and email containers in days — and a team that ships twice a week without breaking the eval.**

This is AIESCU's operator brief: who builds (Teo Deleanu, through AIESCU LLC), how we build, the proof, and how an engagement runs: a free scoping week, then a flat $20k 4-week build sprint and a handover you own. The doc-intel-platform demo in §03 is the kind of MVP I deliver. §09 has the engagement shape, commercial terms, and three ways to continue (or walk away clean) after.

Category: AI document intelligence.

"This artifact is allowed to be honest about what I haven't shipped — without being allowed to bluff what I have."

Headline metrics:

- **99.9%** API SLA (DeFi platform · monthly)
- **<300ms** p95 SWAP QUOTE (production target)
- **<60s** MTTR ON MAIN (CI auto-revert (mosaic-ai))
- **~60s** GITOPS ROLLBACK (ArgoCD (FundAtlas))
- **<30s** ALERT MTTR (streaming rules (rhea+dione))
- **0.93** RECALL@K (llama_pg eval (doc-intel))
- **0.97** FAITHFULNESS (eval-as-CI scorer)
- **<150ms** OPENSEARCH p95 (hybrid query · 3072-dim)
- **10×** COST CUT (Golang notification rewrite)
- **117** PRs / WK PEAK (AI-coded, all reviewed)
- **458** TESTS GREEN (doc-intel-platform proof)
- **3 days** POC → SHIP (spec → demo-able MVP)

## 01 · Scaling document ingestion

The hardest problems on a doc-intelligence roadmap are document problems. Multi-thousand-page PDFs. Scanned images. Email containers. Office files. Archives. Each format has its own failure mode — OCR misroutes, embeddings drift between model versions, queue retries silently double-charge a provider, an "idempotent" job races itself on a redeploy.

These problems need an engineer who has personally shipped these systems — not a doc-AI manager from a distance. Someone who can write the post-mortem AND set the technical bar your team keeps after the handover.

What breaks:

- OCR routing on scanned multi-thousand-page packages
- Format detection (mixed text / image / embedded fonts)
- Splitting + chunking under token budgets
- Queue retry storms after a provider 429
- Idempotency races on redeploys
- Eval regressions silently shipped to prod

What's needed:

- OCR fallback policy + per-stage observability
- Multi-provider router with graceful degradation
- Cassette-everything CI so eval gates can fail loud
- DLQ + replay tooling for ops to recover docs
- SLOs the team actually believes in
- Post-mortems before incidents repeat

## 02 · Thesis

**Player-coach by default.** I still write specs and read every PR. I just don't type the keys — Claude Code does. In my peak week, 117 PRs landed across a production AI platform, all AI-generated, all reviewed by me. That's how I scale a 5–10 person team without becoming the bottleneck.

**Spec-first, eval-as-CI.** Every system I ship starts with a spec that goes through 4–5 review rounds before any code. Then eval gates run on every PR with the delta posted as a comment — if recall@k drops, the merge button is gone.

**Subagent-driven implementation.** I split each project into landing lanes — disjoint file ownership boundaries that let parallel implementer agents (or humans) land code without merge conflicts. Two-stage review per chunk: spec compliance, then code quality.

## 03 · doc-intel-platform — I built the MVP this week

A 15-chunk MVP of a doc-intelligence system: spec-first, subagent-driven, eval-as-CI, two-stage review per chunk. 3 calendar days from empty repo to a 458-pytest / 48-vitest / 3-Playwright-spec green build with multi-provider LLM routing, swappable index backends, OCR fallback, OTel + Prometheus + structlog, and a live /demo flow that streams stage events into the browser via SSE.

The repo is real: github.com/aiescu/doc-intel-platform. Every chunk has its own PR, its own review pass, and a runnable acceptance gate.

Run 1, without me (team builds doc ingestion from scratch): time to RAG-in-prod Q3; 3-5 rework cycles; eval regression found in prod. Sequence: research spike comparing llama_index vs custom RAG (3w); FastAPI + pgvector + naive chunker POC (4w); incident, OCR loop on an 800-page PDF (lost week); refactor of chunker, retry queue, partial idempotency (5w); OpenAI + Anthropic fallback (3w); incident, Gemini 429 cascades through the queue (lost week); manually triggered eval suite (2w); hiring, 2 reqs open with no offers in pipeline (8w).

Run 2, with me (Free scoping week, then a 4-week build sprint on the doc-intel-platform spec): time to RAG-in-prod 1 month; 0 rework cycles; eval regression blocked at PR. Sequence: scoping week (free): customer calls recorded with consent and the deliverable agreed, one ADR per load-bearing decision (chunking · OCR · queue · fallback), a chunked implementation spec with 4-5 review rounds; build sprint (4w): spec chunks shipped subagent-driven, parallel per lane; then automatic two-stage review, eval deltas posted to every PR, runbooks (ingest stuck · provider outage · latency spike), and a 2x/wk eval-gated cadence to main.

The "without me" panel is the sequence of incidents I've watched ship five times across five different teams. The "with me" panel is the sequence of chunks actually merged on github.com/aiescu/doc-intel-platform.

## 04 · Sample portfolio — production systems, with the receipts

Ordered by doc-intelligence relevance. The 2026 production AI contract entry is current, founder-direct, full-time. The 2025 entries are AIESCU LLC customer engagements — hourly-billed, founder-direct. The 2024 entries are from a 9-year platform tenure on an internal IoT analytics + agent stack.

### AI-Native DeFi platform · Multi-model orchestration + MCP toolchain

2026 · Production AI contract · Remote · DeFi / FinTech · Series-B platform (founder name on request behind mutual NDA) · Multi-model LLM · MCP · Agentic pipelines · Azure. Stack: MCP, LiteLLM, Gemini 2.5→3.5, Celery, FastAPI, Datadog, Langfuse, Azure.

Production agentic platform delivering agent-driven financial analytics. Built and own the MCP tool servers (token resolver, asset search, swap quotes, oracle/gas), Redis-backed Celery with adaptive chart derivation, per-period TTL, self-healing cache, and worker scaling. Encrypted API-key storage and multi-env JWT validation across Azure. Led the multi-service Gemini 2.5→3.5 migration with RAG indexing, AI guardrails + PII anonymisation, and eval-scoring pipelines.

SLO: 99.9% API availability · p95 swap-quote < 300 ms · MTTR < 5 min via Datadog APM + Grafana queue-depth + Langfuse prompt traces · Gunicorn × Uvicorn tuning for multi-core utilisation · 117 PRs landed in peak week (player-coach mode). · Reference: founder will take a 30-min call after a mutual NDA — intro via teo.deleanu@gmail.com.

### FundAtlas · DDQ document intelligence + durable workflows

2025 · AIESCU LLC ·PE / VC firms · AI Insights · RAG · Multi-tenant SaaS. Stack: RAG, Temporal, pgvector, EKS, ArgoCD, LGTM.

Multi-tenant SaaS turning PE/VC due-diligence questionnaires into queryable knowledge with Postgres + pgvector RAG. Migrated email ingestion off Celery onto self-hosted Temporal with per-thread keying and sliding-window concurrency to dodge the 2000-pending-activity ceiling.

SLO: ~60s GitOps auto-rollback via ArgoCD · LGTM stack on EKS (Loki / Mimir / Tempo / Grafana Alloy) with S3-backed retention · pgvector RAG recall@5 ≥ 0.92 on DDQ corpora · AWS cost-allocation tag governance via root Terragrunt for per-tenant unit economics.

### mosaic-ai · Editorial AI for healthcare policy

2025 · AIESCU LLC ·Healthcare policy customer · AI Insights · RAG · Collaborative editorial UX. Stack: LangChain 1.x, ChromaDB, SSE, Tiptap+Velt, Prisma AuditLog.

Generates policy memos, articles, infographics, and slide decks with collaborative Tiptap+Velt editing on a RAG corpus. Migrated agent.py to LangChain 1.x with sse-starlette streaming and shipped a CI malware-scan workflow with auto-revert on main after a build-time JS injection incident.

SLO: TTFB < 800 ms on streamed editorial · MTTR < 60 s on main via CI auto-revert · Full Prisma AuditLog (action / status / IP / UA / metadata) on every editorial change · Admin RBAC with last-admin guard.

### Earned Media Tools · AI Composer V2 + embedding ETL at studio scale

2025 · AIESCU LLC ·Major-studio marketing customer · ETL · Per-tenant token accounting · Queue isolation. Stack: Celery, RedBeat, pgvector, Azure, slowapi, Per-tenant cost.

AI Composer + Publicity capture across a major-studio film + streaming portfolio. Five isolated Celery queues (embeddings / ai_relevancy / reports / adhoc / default) with slowapi Redis-backed rate limiting. Per-embedding token_count persisted in DB for tenant cost attribution and per-provider scraping observability tables.

SLO: 99.5% pipeline freshness on weekly release train · Per-tenant unit economics surfaced via token_count rollups · Release Train PR 252 turned recurring broker / Google-Search timeouts into auto-recovery · Build-time JS injection patched in next.config.js.

### Dual-agent recruiting + sales platform with OpenSearch hybrid search

2025 · AIESCU LLC ·Talent + recruitment customer · AI Insights · Agentic · Hybrid search · Document parsing. Stack: LangGraph 0.2, AWS OpenSearch hybrid, PyMuPDF, text-embedding-3-large, Twilio, MS Graph.

Recruiting agent (resume scoring, email screening, interview scheduling) + sales agent (lead extraction from the candidate corpus) backed by AWS OpenSearch hybrid search and text-embedding-3-large (3072 dims). Custom RateLimitHandler with exponential backoff and RFC-7231 Retry-After, plus CATS API exception taxonomy.

OpenSearch hybrid p95 query < 150 ms on the candidate corpus · 3072-dim embeddings · Cost-aware model selection (gpt-5.4-mini capped 200–800 tok for batch matching) · Hourly recommendations cache avoids re-spend · Per-candidate token accounting end-to-end.

### BI connectors + NL→SQL chatbot for tourism analytics

2025 · AIESCU LLC ·Tourism / DMO customer · ETL · BI · Reliability / SRE. Stack: Temporal, LangGraph SQL, ECS circuit-breaker, Redis, structlog.

HubSpot + SimpleView DW + Google Ads consolidation with role-based dashboards and a LangGraph SQL-agent chatbot (gpt-5.5 → gpt-5.4-mini fallback). Caught a failed ECS deploy via deploymentCircuitBreaker; designed the WS scale-out runbook with ALB sticky sessions; Redis → fakeredis fallback so dashboards keep serving during cache outages.

SLO: p99 cached dashboard < 200 ms · cache warmup loop + statement_timeout cut tail latency · ECS deploy circuit-breaker stopped one production rollout from going live · structlog with correlation-ID middleware end-to-end.

### ai-core · Multi-agent orchestrator-specialist analytics platform

2024 · Internal IoT analytics platform · AI Insights · Agentic platform · Text-to-SQL RAG. Stack: LangGraph, Claude Opus 4.8, Vanna RAG, Neo4j, Databricks, Langfuse.

Front-desk orchestrator (Sonnet 4.6) splits multi-domain questions, routes sub-questions to Operational / Analytical specialists (Opus 4.8), then synthesizes — single-domain queries pay zero synthesis tax. Generic-tool pattern replaces 200+ hardcoded tools with one tool + 9k-token compressed catalog.

Tenant-safe text-to-SQL via sqlglot AST application_id injection (zero cross-tenant leak by construction) · AsyncPostgresSaver checkpoints for resumable agent runs · Langfuse-traced end-to-end with per-task cost attribution.

### rhea + dione · Streaming ETL + rules-engine alert system

2024 · Internal IoT data platform · ETL · Alert system · Streaming primitives. Stack: Go, Redis Streams, Elasticsearch ILM, RabbitMQ, Firebolt, Prometheus.

rhea ingests sensor readings into Redis Streams (hot) and Elasticsearch monthly indices (warm) with ILM policies and 90-day gateway-metrics retention. dione's rules engine reads rhea's streams to evaluate basic rules plus DAB / ATA / OACR overrides via the Firebolt streaming framework — replaced legacy CrateDB stack.

SLO: alert MTTR < 30 s end-to-end · sustained ingest throughput at telemetry scale · custom Prometheus metrics on rule-eval latency / failures · rule-lookup query duration · publish failures — surfaced on the on-call dashboard.

### skills · Versioned skills library for MCP-based agents

2024 · Internal agent platform · Agent platform · Skills governance · MCP. Stack: MCP, Skills.md, Apache 2.0, Semver.

Versioned domain expertise (SKILL.md) loaded by MCP, the agent platform, and the PWA chatbot before tool calls. Two-truth data model separates operational ('now') from analytical ('patterns'). Governance: ADOPT / FORK / BUILD classification, quarterly review, semver mandatory, customer-facing vs team-internal segregation.

300+ skills under governance · Apache 2.0 attribution for forked upstream skills · per-skill version pinning surfaces stale dependencies in CI.

## 05 · AI capabilities · evidence matrix

AI engineering theme by theme. Each row is a capability; the receipts are the repos where I shipped it.

### LLM orchestration & multi-model routing

Multi-provider agent graphs, model migrations under live traffic, per-task fallback chains. (LangGraph, LangChain 1.x, LiteLLM, AWS Bedrock, Gemini 2.5→3.5, Claude Opus 4.8, ReAct)

- **chaos/aip**: 44 routers · 55 services · 42 repository classes · 19 DI containers. LangGraph + LangChain over OpenAI / Claude / Gemini. Owned the multi-service Gemini 2.5→3.5 migration.
- **chaos/ai-agents-workflow**: LangGraph research orchestration with tool-execution pipeline across OpenAI / Claude / Vertex AI.
- **internal/ai-core**: Orchestrator-specialist split (Sonnet 4.6 front-desk → Opus 4.8 specialists). Single-domain queries pay zero synthesis tax. AsyncPostgresSaver checkpoints for resumable runs.
- **aiescu/mosaic-ai**: LangChain 1.x migration with sse-starlette streaming over a healthcare-policy editorial corpus.
- **production · AWS Bedrock**: Multi-model routing through Bedrock (Claude · Titan V2 · Cohere Embed v4) on a regulated-data engagement where the embeddings + inference had to stay inside the customer's AWS account.
- **doc-intel-platform**: LLMRouter on LiteLLM with per-task primary + fallback chain. Provider-agnostic by construction; mock-routing override for cassette-only CI.

### Hybrid retrieval · RAG

pgvector + OpenSearch hybrid + Bedrock embeddings — recall measured, not assumed. (pgvector, OpenSearch hybrid, RRF k=60, text-embedding-3-large 3072d, Bedrock Titan V2 / Cohere Embed v4, Tavily, Firecrawl)

- **chaos/ai-search**: Federated Google GenAI / OpenAI / Perplexity / Tavily / Firecrawl. OpenSearch semantic + keyword hybrid. MCP-compliant /semantic_search/execute. Financial-domain indices.
- **chaos/ragindex**: Airflow 15-DAG ingestion into Postgres workmemory schema. Idempotent + schema-versioned migrations. LiteLLM spend logs → queryable AI trace events.
- **chaos/ai-memory-pipelines**: Source-hash-versioned enrichments. 8 consumer views (clustering / routing / eval / Pulse). JSONB prompt+response payloads. Cost-wedge computed at read-time, never stored.
- **production · AWS Bedrock + OpenSearch**: Document-ingestion pipeline using Bedrock embeddings (Titan V2 / Cohere Embed v4) into AWS OpenSearch hybrid index. Chunking + embedding + index sync with multi-tenant retrieval against the ingested corpora. Stays inside the customer AWS account end-to-end (regulated-data engagement).
- **aiescu/FundAtlas**: Postgres + pgvector RAG over PE/VC due-diligence questionnaires. Recall@5 ≥ 0.92 on real DDQ corpora.
- **aiescu/talent-platform**: AWS OpenSearch hybrid + text-embedding-3-large (3072d). p95 hybrid query < 150ms on the candidate corpus.
- **doc-intel-platform**: RRF k=60 hybrid retrieval. llama_pg vs raw_pg swappable IndexBackend. Recall@k 0.93 on the llama backend (eval-as-CI).

### MCP tool servers

Model Context Protocol surface for agent-driven tools — at scale, governed, cached. (FastMCP, MCP SDK 1.26, Redis L2, SHA256 cache keys, Semver, Schema discovery)

- **chaos/aip-mcp-tool-server**: 22 hubs · 850+ tools (CryptoHub 188 · CeFiHub 117 · TradFiHub 103 · RWAHub 67 · SearchHub 34). Semantic reranking 'fast' mode. Per-tool TTL + Redis L2 horizontal cache.
- **internal/mydevices-mcp**: TypeScript MCP server on @modelcontextprotocol/sdk. 42 tools across 5 domains (admin · energy · environmental · health · weather). Scoped user/admin servers for multi-tenant auth (JWT via jose). Skills system loads domain frameworks (energy / energy-frameworks / environmental) before data queries. OTel auto-instrumented, pino-structured logs.
- **chaos/aip-playground-server**: LangGraph runtime + Next.js dev surface for testing workflows with Claude integration.
- **internal/skills**: 300+ versioned MCP skills. ADOPT / FORK / BUILD governance. Semver pinning surfaces stale deps in CI. Customer-facing vs team-internal segregation.

### Reducing LLM calls

Cache, fallback, tool minimization, capped budgets — token bills move under control. (Redis L2, SHA256 keys, Generic-tool catalog, Model fallback, Token caps)

- **chaos/aip-mcp-tool-server**: Redis L2 with SHA256 cache-key hashing + per-tool TTL. Multi-layer L1/L2. 'fast' mode runs a lighter reranker. Thread pool for CPU-bound ops.
- **internal/ai-core**: Generic-tool pattern: 1 tool + 9k-token compressed catalog replaces 200+ hardcoded tools. Slashes context cost per turn.
- **aiescu/talent-platform**: gpt-5.4-mini capped 200–800 tokens for batch matching. Hourly recommendations cache avoids re-spend. Per-candidate token accounting end-to-end.
- **aiescu/tourism**: gpt-5.5 → gpt-5.4-mini fallback on rate-limit. Redis → fakeredis fallback so dashboards keep serving during cache outages.

### External-API integration & agent capture

Provider SDK plumbing in three languages, event webhooks (idempotent, signature-verified, dead-lettered) — plus capture of our own AI workflows, end-to-end. (webhooks, mitmproxy, Rust + Tokio, Go + Gin, Protobuf wire, RFC-7231 Retry-After)

- **chaos/ai-web-capture**: mitmproxy add-on. Provider-specific parsers for Claude / ChatGPT (Wham) / Perplexity. Semantic turn assembly with live freshness watermarks. Optional gated upload.
- **chaos/agent-rs**: Rust relay proxy for Claude Code + Codex CLIs → LiteLLM. SSE reconstruction. Tokio async. In-memory Exchange builder. Two CLI parser paths.
- **chaos/web-capture-api**: Go server (Gin + bun ORM). Protobuf wire (UseProtoNames). 10% trace sampling on high-volume ingest. OTel OTLP/gRPC.
- **chaos/stream-listeners**: asyncio AMQP consumers feeding Work Memory + activity pipelines.
- **aiescu/talent-platform**: Custom RateLimitHandler with exponential backoff + RFC-7231 Retry-After. CATS API exception taxonomy. MS Graph + Twilio integrations.

### Evals + golden metrics

Quality measured on real corpora. Deltas posted to PR. Regressions block merge. (faithfulness, citation_correctness, recall@k, eval-as-CI, llm-judge)

- **doc-intel-platform**: 3 scorers (faithfulness 0.97 · citation 0.95 · recall@k 0.93 on llama_pg). Eval-delta posted to PR comments. Cassette-everything CI — green without provider keys.
- **chaos/ai-scorers**: Shared scoring / evaluation framework. Pydantic-typed rubrics across teams.
- **chaos/ai-memory-pipelines**: workmemory.ai_trace_events_for_eval view exposes success/failure semantics + metadata-only rows for the eval harness.

### Streaming & alerting

SSE to the browser, Redis Streams + ILM warm tiers, rules-engine alerting at telemetry scale. (SSE, Redis Streams, Elasticsearch ILM, RabbitMQ, rules engine, MTTR < 30s)

- **internal/rhea + dione**: rhea ingests sensor readings into Redis Streams (hot) + ES monthly indices (warm) with 90-day ILM. dione rules engine with DAB / ATA / OACR overrides via Firebolt. End-to-end MTTR < 30s. Replaced legacy CrateDB stack.
- **chaos/stream-listeners**: asyncio event consumers for Work Memory + activity pipelines.
- **doc-intel-platform**: SSE progress on /demo · Redis doc:{id}:progress + Postgres events · /runs/{run_id} replay surface.
- **aiescu/mosaic-ai**: sse-starlette streaming on editorial responses. TTFB < 800ms on streamed surface.

### Scalability & SRE for AI workloads

Queue isolation, durable workflows, auto-rollback, capacity tuning — the substrate the AI runs on. (Celery isolation, Temporal, ARQ, EKS, ArgoCD ~60s rollback, OTel)

- **chaos/chaos-ai-utils**: Shared OTel + AWS Firehose substrate. Pre-configured meter / tracer providers with batch processors. Retry with exponential backoff. Pydantic settings.
- **aiescu/FundAtlas**: Self-hosted Temporal with sliding-window concurrency to dodge the 2000-pending-activity ceiling. ArgoCD GitOps ~60s auto-rollback. LGTM stack on EKS with S3-backed retention.
- **aiescu/EMT**: 5 isolated Celery queues (embeddings / ai_relevancy / reports / adhoc / default). slowapi Redis-backed rate limiting. Per-tenant token_count rollups for unit economics.
- **aiescu/tourism**: ECS deploymentCircuitBreaker caught a failed prod rollout before going live. ALB sticky sessions for WS scale-out. fakeredis fallback during cache outages.
- **doc-intel-platform**: ARQ queue with heartbeat-supervised workers (worker exits non-zero on dead heartbeat). MAX_CHUNKS_PER_DOC cap → TerminalStageError lands in DLQ.

### Desktop apps & voice — Tauri / Electron / speech-to-text

Cross-platform desktop clients and real-time voice pipelines wired to the same agent backends. (Tauri, Electron, Rust, TypeScript, Speech-to-text, Realtime streaming)

- **aiescu/geek-bye**: Electron live-work assistant shipped as two products (pavleur.com · geekbye.com). Mic + system-audio capture · 16kHz mono S16LE in ~128ms chunks · single reused AudioContext (Chromium ~6/page cap) · 2.2MB/45s reconnect replay buffer · WS→HTTP/SSE transport fallback.
- **aiescu/geek-bye-backend**: Fly.io STT proxy routing Deepgram Nova-3 (EN, 50 slots) vs ElevenLabs Scribe v2 (other languages, 6 slots). Silent-PCM keepalive vs ~16s upstream idle-close. Churn detection (3 reconnects/2min) auto-swaps providers. Voice-agent rehearsal via Deepgram/ElevenLabs conversational APIs.

### Rust & systems infrastructure

Hot-path services and CLI tooling in Rust and Go where Python's overhead is the bottleneck. (Rust, Tokio, Go + Gin, Protobuf, gRPC, SSE)

- **chaos/agent-rs**: Rust relay proxy for Claude Code + Codex CLIs → LiteLLM. SSE reconstruction. Tokio async. In-memory Exchange builder. Two CLI parser paths.
- **chaos/web-capture-api**: Go server (Gin + bun ORM). Protobuf wire (UseProtoNames). 10% trace sampling on high-volume ingest. OTel OTLP/gRPC.
- **receipts on request**: Additional Rust infrastructure work under NDA — walkthrough on the scoping call.

Stack at a glance:

- **Models & gateway**: Claude Opus 4.8 / Sonnet 4.6; OpenAI GPT-5.5 / GPT-5.4 mini; Gemini 2.5→3.5 · Vertex AI; AWS Bedrock · Titan V2 / Cohere Embed v4; LiteLLM — unified gateway + per-task fallback
- **Agents & retrieval**: LangGraph · LangChain 1.x; MCP tool servers (FastMCP); pgvector · OpenSearch hybrid · RRF k=60; text-embedding-3-large (3072d); eval-as-CI — faithfulness · citation · recall@k
- **Backend & data**: Python · TypeScript · Go · Rust; FastAPI · Celery · Temporal · ARQ; PostgreSQL · Redis · Elasticsearch; Snowflake · Databricks · Neo4j
- **Infra & observability**: AWS · EKS / ECS · ArgoCD GitOps; OpenTelemetry · Prometheus · Langfuse; Datadog · Grafana LGTM stack; Docker · Terragrunt

## 06 · What to check before you hand over an AI build — with the evidence

What should be true of whoever builds your production AI. Here's that checklist, with the receipt for each row.

- **15+ yrs engineering + hands-on tech leadership**: 15+ yrs total. Lead at Comodo / MiOS / forfone (2008-2013, 4-16 engineers across 3 product lines). Lead at an IoT platform org (2014-2023, 9 yrs owning platform + cloud infra). Production AI contract since 2025.
- **Built & managed teams, hired seniors, set up review / on-call / shipping**: 9-year platform tenure with team build-out from MVP forward. CV §6 names it: agile delivery, code review, observability-first design, incident response, on-call rotations, mentoring 4-16-engineer teams.
- **Deep production-grade document processing experience**: doc-intel-platform (Jun 2026): 6-stage pipeline + OCR fallback + DLQ + idempotency. FundAtlas DDQ RAG with Temporal durability. Talent-platform recruitment agent with PyMuPDF resume parsing at OpenSearch hybrid-search scale.
- **Speak in detail about failure modes / monitoring / tradeoffs**: 5 ADRs + 4-angle post-MVP audit (security · prod-readiness · perf · contract) on doc-intel-platform. Studio-marketing queue isolation + per-tenant token accounting. Legacy rabbot → amqplib migration with bounded concurrency.
- **Player / coach (split dialed where the team needs it)**: 9 years in that mode on the IoT platform — dialed coach-heavy when the team was humming, player-heavy when prod was on fire. Currently still hands-on across 5+ AIESCU LLC customers AND the production AI contract AND running my own engineering process on doc-intel-platform — that's the explicit player end of the dial.
- **Multi-model orchestration (OpenAI / Gemini / Anthropic)**: ai-core LangGraph orchestrator-specialist with Claude Opus 4.8. doc-intel-platform LLMRouter with all 4 providers + per-task fallback chain. Gemini 2.5 → 3.5 multi-service migration on the production DeFi platform.
- **RAG retrieval quality + eval infrastructure**: doc-intel-platform: hybrid retrieval (RRF k=60) + 3 scorers + eval-as-CI with delta posted to PR (faithfulness 0.97 · citation 0.95 · recall@k 0.93 on the llama_pg backend). mosaic-ai LangChain 1.x RAG over policy docs. Talent platform on AWS OpenSearch hybrid search.
- **Engineering hiring + bar setting**: Comodo / MiOS / forfone: hired and led 4-16 engineers across 3 product lines. IoT-platform org: built the team alongside the platform from MVP onward. Set the spec-review and chunked-shipping bar for current contracts.
- **Roadmap + ~2x / week shipping cadence**: IoT-platform tenure: 'delivered features daily.' Current DeFi contract: 117 PRs in peak week. doc-intel-platform: 15 chunks merged in 3 days via subagent-driven flow.
- **Production reliability — on-call, SLOs, incident response**: Studio-marketing platform: turned recurring incidents into auto-recovery (Release Train PR 252). Tourism-BI platform: caught failed ECS deploy via deploymentCircuitBreaker. doc-intel-platform: docs/slos.md + multi-component /healthz + heartbeat supervision exits worker on death so docker restarts cleanly.
- **SRE coverage · SLA targets · MTTR**: DeFi platform: 99.9% API availability · p95 swap-quote < 300 ms · MTTR < 5 min via Datadog APM + Langfuse traces. FundAtlas: ~60 s GitOps auto-rollback via ArgoCD. mosaic-ai: < 60 s MTTR on main via CI auto-revert. IoT platform: alert MTTR < 30 s end-to-end. doc-intel-platform: SLOs documented; healthz with db / redis / vector_meta / worker_registered checks.
- **RAG quality measured — recall@k, faithfulness, citation correctness**: doc-intel-platform: 3 scorers (faithfulness · citation_correctness · recall_at_k) running on every PR with delta posted as a comment; merge blocked on regression. FundAtlas: pgvector recall@5 ≥ 0.92 on DDQ corpora. Talent platform: OpenSearch hybrid p95 < 150 ms.
- **Unit economics · cost per inference · per-tenant attribution**: Studio-marketing: per-embedding token_count persisted in DB for tenant cost rollups. Talent platform: cost-aware model selection (gpt-5.4-mini capped 200–800 tok for batch matching) + hourly recommendations cache. IoT-platform Golang notification rewrite: 10× server-cost reduction + materially lower SMS spend. ArgoCD + Terragrunt cost-allocation tag governance on AWS.
- **Observability stack (traces · metrics · logs · prompts)**: Datadog APM + Grafana dashboards + Azure App Insights + Langfuse + OpenTelemetry across the DeFi platform. FundAtlas: full LGTM stack (Loki + Mimir + Tempo + Grafana Alloy) on EKS with S3 retention. Tourism-BI: structlog + correlation-ID middleware end-to-end. doc-intel-platform: OTel auto-instrumentation (FastAPI + psycopg + httpx) + Prometheus stage metrics + structlog.
- **Cross-functional / founder-direct**: All AIESCU LLC customers are founder-direct by definition. IoT-platform tenure: 9-year bridge between platform team and customer-facing teams. DeFi contract reports directly to founder.

## 08 · How I work

- **Player ⇄ coach — dial it where you need it (20/80 ⇄ 80/20)**: 20% coach / 80% player when your team needs me writing specs and shipping code alongside them; 80% coach / 20% player when your team is humming and needs me to set the bar, review, and clear blockers. We pick the ratio explicitly in the scoping week.
- **Spec-first review rounds (5×)**: every non-trivial spec gets 4–5 review passes before code.
- **Shipping cadence (2x/wk)**: twice-a-week to main, behind eval-as-CI gates. Cadence is preserved by tooling, not by heroics.
- **AI-coded (100%)**: every line since 2025 is Claude Code generated. Peak week: 117 PRs. The discipline is in spec quality and review depth, not keystrokes.
- **Subagent review (2-stage)**: per chunk, spec-compliance reviewer first, code-quality reviewer second.
- **Untested merges (0)**: eval-as-CI + Playwright e2e + pytest + vitest + ruff + pyright. PR is red until all green, then still red until reviewed.
- **Codex + Claude Code (2 CLIs)**: Codex for fast, scoped iteration; Claude Code for spec-driven multi-chunk work. Both run in subagent mode against the same spec.
- **Manager mode (5→1)**: I manage AI agents the way I'd manage humans. A team of one delivers what a 5-person team would, because the bottleneck is spec quality and review depth, not keystrokes.
- **Skills as managed deps (300+)**: versioned skill library loaded by both CLIs before tool calls. ADOPT / FORK / BUILD governance. Semver pinning surfaces stale deps in CI.
- **Workstreams · landing lanes**: disjoint file-ownership boundaries so parallel agents (or humans) land code without merge conflicts. doc-intel-platform was 14 chunks across 5 lanes (ingest · index · retrieve · observe · eval).
- **Monolith for LLMs · μ-services for humans**: LLMs work fast against a monolith; humans scale on microservices. The art is picking the architecture that lets a small fleet of LLMs work fast without forcing humans into a topology they can't orchestrate.
- **Virtual office · async + documented (24/7)**: async-first, Looms instead of standups, customer calls recorded with consent, every decision documented.

## 09 · How to engage — free scoping week, then a 4-week build sprint

Remote, ET-timezone delivery via AIESCU LLC. A free 30-minute scoping call, a free scoping week, then a flat $20k 4-week build sprint and a handover. You keep the code, ADRs, spec, eval suite and skills — no lock-in, no retainer. Larger scopes get a fixed multi-sprint quote.

The differentiator: customer-call recordings with consent (Granola / Fireflies / Otter) become the source of truth for the spec. Founders rarely get a documented research trail of their own customers' pain — the scoping week produces one.

- **Call**: Free 30-minute scoping call, direct with Teo: the problem, the data, and whether a sprint fits.
- **Scope**: Free scoping week — discovery + spec. Customer calls recorded with consent, a deep-dive on the current pipeline + ops, and one ADR per architecturally-load-bearing decision (chunking · OCR routing · queue topology · multi-provider fallback · eval gates). We agree the deliverable and write the chunked implementation spec; you review it chunk by chunk before the build starts.
- **Weeks 1–4**: The 4-week build sprint: the spec shipped chunk by chunk to production, subagent-driven, behind eval-as-CI gates. doc-intel-platform (§03) is the shape of that MVP.
- **Handover**: Code, ADRs, spec, eval suite, and the specialized skills transferred to you. No lock-in, no retainer.

Commercial terms (redlineable):

- **Price**: The scoping call and the scoping week are free. The build sprint is a fixed fee of $20k per 4-week build sprint — flat, no hourly billing. 50% on signature / 50% on Week-4 handover. 4 weeks of my calendar held. Larger scopes get a fixed multi-sprint quote.
- **Kill-switch · flat 25%**: Either side can stop after the first week of the build sprint. You pay a flat 25% of the fee ($5k); the rest is refunded or not charged. You keep the research corpus.
- **IP ownership**: you keep the research corpus · ADRs · implementation spec · code · eval suite · skills. I retain reusable patterns from doc-intel-platform (already in the open repo) for future engagements.
- **Logistics**: Contracted through AIESCU LLC. Mutual NDA before kickoff. Customer calls recorded with consent (Granola / Fireflies / Otter). Weekly founder check-in. Termination: 5 business days' notice either side.

After the handover — three ways forward:

- **A. Fractional CTO / advisor**: 1–2 days/week at the market day rate for senior fractional engagements, ET hours, quarterly on-site visit. Continuity on the spec without a full-time hire.
- **B. Next sprint**: keep building against the spec with another fixed-fee sprint, or a fixed multi-sprint quote for a larger scope. No retainer, no headcount commitment.
- **C. Clean walk-away**: you keep the code, the ADRs, the spec, the eval suite, and the skills. I get a paid sprint and a real reference.

## 10 · Risks, lessons, open questions

- **Named risks I'll watch from the scoping week**: eval corpus availability (mitigation: budget corpus curation into the spec); OCR vendor lock-in (mitigation: spec the abstraction in a scoping-week ADR).
- **Vendor and architecture decisions I've walked back**: Celery + RedBeat to RabbitMQ + bounded concurrency on the DeFi platform after 3 weeks of retry-storm pain; Temporal on FundAtlas, overruled at first, migrated 6 months later with the 2000-pending-activity ceiling as evidence; a 200-tool MCP server replaced with a single generic tool + 9k-token catalog (lost a month).
- **What I'd improve building a knowledge / second-brain system today**: retrieval quality over model size; capture at the source; agentic, not just lookup; provenance + freshness on every answer; evals as a first-class loop.
- **What I'd need from you**: a real problem, access to (or a sample of) your data, and the free scoping call. During the engagement: a weekly founder check-in (recorded), comfort with spec-first meaning the scoping week settles the spec before any build code, and read access to the customer-success Slack.
- **What I'm still figuring out**: whether AI-coded shipping scales to a 10-person team or breaks at 5; skills-library investment vs just-in-time agent prompts as context windows grow; where MCP tool servers stop being the right primitive and direct-API calls win again.

## More

- Done-for-you engagements: https://aiescu.com/dfy.md
- Full site index: https://aiescu.com/llms.txt