Skip to content

One Developer. Four Months.
A Medical AI Platform.

A solo developer used FlowForge to build DELPHOS — a production-grade medical AI platform serving a live clinic network in Brazil. Fourteen clinical domains. Five on-premise AI models. Brazilian healthcare compliance (LGPD, ANS, TISS, TUSS, CBHPM). Every billed hour traceable to a ticket.

1 Developer
4 Months
14 Domains
5 AI Models
80%+ Test Coverage

By the numbers

1.52M Lines of code authored
97/100 Documentation quality seal
5 On-premise AI models in production
14 Clinical domains covered
11 Parallel workers per developer

Without FlowForge vs. with FlowForge

  • Hire 6–8 developers for 18+ months
  • $800K–$1.2M in salary before a single patient is seen
  • Manual code reviews — quality is whoever has time
  • Context loss every time a developer leaves or a session ends
  • No audit trail — "how long did that actually take?" is unanswerable
  • AI tools write unreviewed code directly to main
  • 1 developer + AI workforce in 4 months
  • Fraction of the cost — every hour billed, no idle time charged
  • Automated quality gates on every commit — complexity, coverage, docs
  • Session handoffs preserve context across conversations and developers
  • Every minute maps to a ticket, a session, and an event log
  • 33 specialist agents — each domain reviewed by its specialist

The technical stack

DELPHOS runs entirely on-premise on K.I.I.T. No data leaves the clinic network.

Language models

  • LLM DELPHOS — primary NLP and orchestration. Handles prescription processing, clinical note parsing, and multi-domain routing.
  • Reasoning DELPHOS — structured reasoning. Drug interaction analysis, dosage verification, differential diagnosis support.

Retrieval & embeddings

  • Embeddings DELPHOS — multilingual embeddings for Portuguese clinical text. Semantic search across patient history and clinical guidelines.
  • Cross-encoder reranker — precision layer over vector retrieval to surface the most clinically relevant passages.

Multimodal

  • Transcriptor DELPHOS — clinical speech-to-text in Brazilian Portuguese. Real-time transcription during consultations.
  • DELPHOS Inteligência Documental — document intelligence. Extracts structured data from scanned CBHPM forms, lab results, and referrals.

Infrastructure

  • vLLM — inference serving with continuous batching. Manages GPU allocation across concurrent clinical sessions.
  • K.I.I.T. — primary on-premise server. K.A.R.R.-01 on standby as secondary.
DELPHOS architecture on K.I.I.T. — vLLM router fans inference across five DELPHOS models, backed by pgvector retrieval and Mnesis EHR. Three-tier schematic. Top tier: clinic interfaces (Web EHR, Mobile App, Consultation Room) send queries to the AI inference tier. Middle tier: vLLM router routes to five DELPHOS models — LLM DELPHOS, Reasoning DELPHOS, Transcriptor DELPHOS, DELPHOS Inteligência Documental, and Embeddings DELPHOS — running on continuous batching. Bottom tier: infrastructure layer including the K.I.I.T. on-premise server, pgvector retrieval database, Mnesis EHR with calendar, Agenda Inteligente scheduling, and K.A.R.R.-01 as failover standby. CLINIC INTERFACES Web EHR clinician console Mobile App on-call access Consultation Room live transcription AI INFERENCE vLLM continuous batching LLM DELPHOS primary NLP orchestrator Reasoning DELPHOS drug interaction dosage check Transcriptor DELPHOS PT-BR speech-to-text DELPHOS Inteligência Documental OCR / extraction Embeddings DELPHOS multilingual retrieval INFRASTRUCTURE K.I.I.T. on-prem server pgvector vector retrieval semantic search Mnesis EHR calendar patient records Agenda Inteligente AI scheduling engine K.A.R.R.-01 failover standby

Designed for Brazilian healthcare compliance

Clinical AI in Brazil operates under a layered regulatory framework. DELPHOS was designed from day one to respect every layer.

Physician autonomy (Lei 12.842/2013)

The law reserves clinical decision-making exclusively to physicians. DELPHOS alerts, informs, and presents evidence — it never blocks or prevents a prescription. The design principle: a digital system cannot be more restrictive than paper. Every interaction is advisory, not gatekeeping.

ANS, TISS, TUSS, CBHPM coverage

Billing, procedure coding, and authorization flows are mapped to the ANS standardized tables. TISS interchange formats, TUSS procedure codes, and CBHPM fee schedules are ingested and kept current. The AI reasons over live regulatory data, not static snapshots.

LGPD data residency

All patient data is processed on-premise at the clinic network. Zero data leaves Brazilian jurisdiction. The on-prem vLLM stack was chosen specifically to satisfy LGPD requirements without sacrificing model capability.

Portaria 344/98 — prescriptions

Controlled-substance prescription rules (format, quantity, classification) are enforced at the UI layer. Clinical judgement on drug selection remains with the physician — consistent with the advisory-not-blocking design principle throughout the system.

1.52M lines — the methodology

Every claim on our landing page is empirically verifiable. The 1.52M figure is what was built with FlowForge — the DELPHOS family of repositories plus K.I.I.T. infrastructure. FlowForge itself (the tool) is shown separately and excluded from the headline count. Full breakdown below.

Repository / Surface Source Code Docs + Config Total
DELPHOS main 988,313 210,958 1,199,271
DELPHOS frontend (test) 255,995 6,051 262,046
DELPHOS iOS PatientApp 41 6,779 6,820
llama-vision-api 4,378 5,352 9,730
/srv/caddy infrastructure 715 — 715
/srv/backup scripts 77,800 — 77,800
DELPHOS family + K.I.I.T. infra (what was built) 1,327,242 229,140 1,556,382
FlowForge (tool — excluded from count) 376,580 469,245 845,825
ff_backend (tool — excluded from count) 0 33 33
All surfaces combined (for reference only) 1,703,822 698,418 2,402,240

Why 1.52M and not 2.4M? The 1.52M figure counts only what was built with FlowForge — the DELPHOS family of repositories (DELPHOS main, DELPHOS frontend sandbox, DELPHOS iOS PatientApp, llama-vision-api) plus K.I.I.T. infrastructure scripts. FlowForge itself is the tool that produced this output; it is intentionally excluded from the count. The combined total across all surfaces (including FlowForge) is 2.4M lines — shown above for transparency, not as the headline figure. Additional excluded surfaces: K.I.I.T. system configurations outside /srv/caddy and /srv/backup, K.A.R.R.-01 setup, Ansible playbooks, integration glue scripts, 22 Docker container configurations, reference data imports (CBHPM/CMED), and the Mnesis EHR front-end (third-party client scope — excluded to avoid attribution ambiguity).

What counts? Source code (Python, Go, TypeScript, Bash, Swift, SQL, et al.), authored documentation, and configuration files. Vendored dependencies, generated files, and third-party contributions are excluded.

Founder perspective

"I've been writing software since I was fourteen years old. For thirty-six years I watched the same problems repeat: context lost between sessions, time billed without accountability, quality degrading under delivery pressure. FlowForge is the answer I spent thirty-six years wishing someone else would build."

— Alexandre Reis Corrêa Cruz, Founder, FlowForge & DELPHOS

Build your software the same way.

One developer. The right tooling. A system that enforces quality at every commit, tracks every billable minute, and preserves context across every session. That is FlowForge.

Start Building See Pricing

Installs as a local binary — FlowForge never touches your source code. Upgrade to Solo ($29/mo) for the TUI Controller and parallel workers.