← Applied AI Engineering
Technical deep-dive · AgentForge

The infrastructure, end to end.

AgentForge is a production multi-agent operating system that runs four of my business units unattended — sourcing, diligence, modeling, outreach, and reporting. This is the full stack: the agents, frameworks, router, tools, memory, evals, and automations. Built and maintained solo.

35K+ LOC Python20 agents · 6 teams9-tier router26 production APIs
0K+
Lines of production Python
0
Specialized agents, in 6 teams
0
Tool modules (44+ tool definitions)
0%
Cost reduction via prompt caching
0
Model-router tiers, 8 providers
0
Production API integrations
0
Claude Code skills in production
0
Business units run unattended
Architecture

One request, seven layers.

Every interaction flows the same path — capture, orchestrate, route, execute, remember, score, report — each layer independently observable and fault-tolerant.

01

Intake & routing

Universal capture from Telegram, Fireflies, and HTTP → semantic dedup (Voyage embeddings, 0.93 cosine) → structured classification → auto-route at ≥0.85 confidence, else a human-review queue.

02

Orchestration

A CEO-supervisor agent decomposes the request into a MECE task DAG, then dispatches to the right team — wrapped in a 9-layer middleware stack (caching, retries, circuit breakers, cost tracking, call limits).

03

Agent teams

Six specialized teams (20 agents) execute — consulting & PE, coding, strategy, opportunity discovery, LBO modeling, introducer enrichment — with parallelized independent sub-tasks.

04

Model router

A 9-tier cost-optimized router picks the cheapest model that meets the bar (Groq → Gemini → MiniMax → Claude), with automatic circuit-breaker fallback per tier.

05

Tools

44+ tool definitions across 56 modules — research, enrichment, financial modeling, outreach, dev, media — invoked under per-run tool-call and cost limits.

06

Memory & knowledge

Dual persistence — Mem0 semantic memory + a Graphiti/FalkorDB temporal knowledge graph — plus hybrid RAG (Qdrant vectors + BM25 + cross-encoder rerank) over ingested documents.

07

Evals, observability & report

Per-section QA gates, an independent test agent, and full Langfuse tracing (cost, latency, tokens) on every run — then results report back to the operator.

The stack

Modern agentic infrastructure.

The frameworks and services the system is actually built on — not a wish-list.

Frameworks & runtime

LangChain 0.3LangGraph 1.1Claude Agent SDKAnthropic SDKFastAPINext.js 16Python 3.13asyncioPM2launchd

Memory, knowledge & data

Mem0 / LangMemGraphitiFalkorDBQdrantVoyage AI embeddingsDoclingSupabase pgvector ×2MegaRAG hybrid

Models & observability

Claude Opus · SonnetGemini FlashGroq Llama 4MiniMax M2.7Perplexity sonarLangfusestructlog
Agent teams

Twenty agents. Six teams. One operator.

11

Consulting & PE

Planner (Opus) · sub-agents (Sonnet) · QA gates

Does: URL extraction → research → market intel → company profile → deal score → PE value-creation plan → personalized approach letter. Benefit: per-section quality gates (8+/10 to pass) and a parallelized MECE DAG — a full diligence pipeline for $2–5.

6

Coding

Product → Architect → Code → Test → Review → Heal

Does: ships features end-to-end; the test agent reads the PRD (not the code) to prevent coverage theater; the Review/Heal loop runs tsc + biome and self-corrects up to 5 iterations. Benefit: coherent, validated code at MiniMax cost.

4

Opportunity discovery

Signal → Pattern → Synthesize → Report

Does: scans 40+ verticals for weak signals, finds cross-domain patterns, scores on 5 dimensions (readiness, competition, buildability, scalability, defensibility). Benefit: spots trends before mainstream awareness — ~$1.50 a scan.

12

LBO financial modeling

12-node graph · document extraction · SBA gates

Does: deal intake → Docling/Opus extraction → quality-of-earnings → SBA prescreen gate → formula-driven 17-tab 3-statement Excel → 3-scenario sensitivity. Benefit: IC-grade models in under five minutes; 50+ QA-validated.

7

Introducer enrichment

7-stage funnel · provider race · verify-before-send

Does: discover → enrich (Firecrawl/Apollo) → paid-tier deep → web-search → validate → recover → sync. Benefit: verified 56% email coverage — 4,100+ deliverable contacts from a 7,000+-firm cohort.

3

Intake routing

Capture · dedup · classify · route

Does: turns voice, photos, and notes into deduplicated, classified, confidence-scored work items. Benefit: nothing is dropped, nothing is processed twice, and only the uncertain reaches a human.

Model router

The cheapest model that clears the bar.

A 9-tier router across 8 providers, each tier with a circuit-breaker fallback. Caching cuts repeat-prompt cost ~90% (~$270/mo verified).

T0Claude Sonnet 4.6fallback · Gemini FlashSupervisor routing across 44 tools, with auto-caching
T1Groq Llama 4 Scoutfallback · Gemini Flash-LiteClassification, intent detection, fast routing
T2Gemini Flash-Litefallback · Groq ScoutExtraction — Mem0 facts, RAG chunks, parsing
T3Gemini Flashfallback · Groq GPT-OSSMulti-step reasoning, strategy analysis
T4MiniMax M2.7fallback · Gemini FlashCoding team & consulting drafts (Anthropic-compatible)
T5Gemini Flash + Searchfallback · Groq AdvancedGrounded research with citations
T6Claude Sonnet 4.6fallback · Gemini FlashHigh-quality consulting sub-agents
T7Claude Opus 4.7fallback · Sonnet 4.6Client-facing memos, CIM analysis, deep thinking
Tool catalog

44+ tools the agents actually call.

Research & intelligence

8 tools

  • Perplexity — grounded web research + citations
  • CellCog — multi-source synthesis
  • Exa — semantic company search
  • Tavily / SerpAPI — fast web + SERP
Company & lead enrichment

12 tools

  • Apollo — org + people, paid-tier reveal
  • LeadMagic / Datagma — email find + validate
  • Firecrawl — scrape + typed extraction
  • Google Places · PDL — discovery + scoring
Financial & tax

5 tools

  • LBO generator — 17-tab formula-driven Excel
  • Financial extraction — Opus over CIM/K-1
  • Tax engine — W-2 + S-Corp, 25 strategies
  • EBITDA / revenue — sizing for screening
Memory & knowledge

3 tools

  • Mem0 — long-term semantic facts
  • Knowledge graph — temporal entities
  • MegaRAG — hybrid retrieval + citations
Comms & outreach

4 tools

  • Microsoft Graph — email from own domain
  • Smartlead — campaign sequencing
  • Resend — transactional mail
  • Twilio — SMS / voice (optional)
Dev & infrastructure

11 tools

  • GitHub — clone, commit, push, deploy
  • Vercel — auto-deploy on main
  • File + bash — via Claude Agent SDK
  • Supabase · Qdrant · Langfuse — data + traces
Pipelines

Where it does real work.

Verified, in production — not demos.

A

Industry research → lead discovery

launchd · every 6h

A daily AI thesis → 20 deep sub-niches → a Google Places / Exa / Apollo / Firecrawl discovery waterfall. Verified: 631 leads from one thesis in 16 min 33 s.

B

Introducer enrichment

7-stage · fire-and-forget

Discover → enrich → paid-deep → web → verify → recover → sync, with provider-error isolation. Verified: 4,100+ deliverable emails (56% coverage of a 7,000+-firm cohort).

C

LBO modeling

12-node graph · 2 hard gates

Documents in → quality-of-earnings → SBA prescreen → 3-statement model → sensitivity → cross-check. Verified: 50+ IC-grade models, formula-driven.

D

Coding team

5 agents + heal loop

Feature request → PRD → architecture → code → independent tests → tsc/biome review → self-heal. Benefit: validated features at draft-model cost.

Reliability

9-layer middleware on every agent.

01–03

Summarization · tool retry · tool-call limit — survives long runs and transient failures; caps runaway chains at 30 calls.

04–06

Model-call limit · model retry · prompt caching — caps per-run cost, auto-recovers 429/500/503, and caches for ~90% savings.

07–09

Filesystem · sub-agent spawn · cost tracking — safe I/O, asyncio parallelism, and per-call cost logged to Langfuse.

Evals & observability

Nothing ships unscored.

QA

Per-section quality gates — each artifact scored 1–10, 8+ to pass, max 2 rewrites.

Test

Independent test agent — reads the PRD, not the code, so it can't game coverage.

Dedup

Semantic dedup — Voyage embeddings, 0.93 cosine, before anything is stored.

Trace

Langfuse — cost, latency, tokens, and full traces on every run.

Skills

19 Claude Code skills — reusable, versioned.

One-off agent workflows turned into products the system can invoke on demand.

acquisition-researchcreating-financial-modelssba-acquisition-analysismanagement-consultingstrategy-scangtm-onboardingparadosi-outreachpremium-site-builderfrontend-designvideo-to-websitenano-banana-2skp-rendervisualizationsexcalidraw-diagramtax-forgejob-forgedomain-warmupxlsxzero-defect-validator
In production

Four business units, run by the same system.

Search fund

Paradosi Partners

  • Sourcing, enrichment, scoring, and outreach for a lower-middle-market industrial search — 18,000+ leads and 7,000+ referral relationships under management.
AI consulting

ClearForge.AI

  • Strategy factory + SightForge sales intelligence + LBO modeling delivered as client engagements.
Real estate

Investment scanner

  • Auction and foreclosure scanners with per-property investor diligence and automated deep research.
The ask

The repo is private — it runs my businesses. I'll screen-share it live.