L1 · What should agents do?
Venkat AI Platform
Multi-agent orchestration OS
Tradeoff: LangGraph stateful graphs over linear chains — enterprise workflows need checkpoints and HITL.
Metrics loading or API waking…
6-spine review path (5 live + ADR hub) · 17 repos
AI architect portfolio — production systems you can click, inspect, and fork.
Applied AI engineering portfolio spanning agent governance, multi-agent orchestration, enterprise RAG, and AgentOps. Every project ships with its architecture diagram, a live platform, and the open-source repository on one card.
Start Here (5 min)
Technical review path for engineering panels
AegisAI → VAP → Enterprise RAG → ModelForge → Content Factory → ADRs (Demo vs Strict labeled). Full checklist →
Eagle-eye org stack
Ten architecture questions — each with a live demo and production metrics
Click into any layer. Metrics pull from public /ops/metrics on Render — architecture decisions backed by running systems, not slide decks.
L2 · What are agents allowed to do?
AegisAI
Gateway, policy, HITL, audit
Tradeoff: Control plane over agent builder — enterprises need governance across fleets, not another chat UI.
Metrics loading or API waking…
L3 · How do we route and quota model calls?
aegis-llm-gateway
LLM gateway plane — OpenAI-shaped proxy, stub/BYOK
Demo · metrics on hosted API when deployed
L4 · How do we similarity-cache completions?
aegis-semantic-cache
Semantic cache plane — tenant-scoped embed + lookup
Demo · metrics on hosted API when deployed
L5 · What knowledge can they use?
Enterprise RAG Platform
Access-aware governed RAG
Tradeoff: Authorization before ranking — vector DB is implementation; access control is architecture.
Metrics loading or API waking…
L6 · Which weights, where, with what proof?
ModelForge
Model Plane — SLM · PEFT · CUDA vLLM · LLMOps
Tradeoff: One ModelForge flagship beats burying PEFT/vLLM in a teaching drawer (ADR-034).
Demo · metrics on hosted API when deployed
L7 · How do we adapt models to domain format?
DomainForge
RAG + QLoRA triage pipeline (feeds ModelForge)
Tradeoff: RAG for facts, LoRA for schema — never memorize policies in weights.
Metrics loading or API waking…
L8 · How do we run real-time voice triage?
VoiceForge
ASR → LLM → TTS with latency budgets
Tradeoff: Browser ASR on Render; server Whisper opt-in on GPU hosts.
Metrics loading or API waking…
L9 · How do we operate agent fleets?
AegisLoop AgentOps
Missions, traces, eval gates
Tradeoff: Missions over monolithic prompts — specialist handoffs with observable traces.
Metrics loading or API waking…
L10 · What do they produce?
AI Content Factory
Governed content publish pipeline
Tradeoff: AegisAI gateway blocks publish until policy allows — side effects never bypass governance.
Metrics loading or API waking…
L11 · How do agents improve?
LoopForge
Loop harness · RAG tuning · memory
Tradeoff: Harness + eval loops over bigger prompts — ODAEU tunes RAG on failure.
Metrics loading or API waking…
L12 · How do we serve LLMs at scale?
vLLM Architecture Lab
PagedAttention · continuous batching · KV budget
Tradeoff: Educational simulator before GPU spend — teams learn inference architecture without H100 budget.
Demo · metrics on hosted API when deployed
L13 · How do we answer multimodal asks with the right models?
OmniForge
Multi-agent multi-LLM routing lab
Tradeoff: Self-contained monorepo — no sibling runtime deps; routing is architecture, not a dropdown.
Demo · metrics on hosted API when deployed
L14 · How do we scan overnight signal?
Sentinel Brief
Governed intelligence brief · email
Tradeoff: Autonomous inner loop, governed outer loop — only email send goes through gateway.
Metrics loading or API waking…
Agent FinOps
Usage metering + budget breach detection — gates LLM dispatch across the fleet.
Metrics loading…
Governed AI Reference Stack
14 architecture questions. 13 production answers.
Each layer answers a question enterprise agent programs must resolve before production.
2
What are agents allowed to do?
Gateway, policy, HITL, audit
aegisai-enterprise-agent-platform
3
How do we route and quota model calls?
LLM gateway plane — OpenAI-shaped proxy, stub/BYOK
aegis-llm-gateway
4
How do we similarity-cache completions?
Semantic cache plane — tenant-scoped embed + lookup
aegis-semantic-cache
6
Which weights, where, with what proof?
Model Plane — SLM · PEFT · CUDA vLLM · LLMOps
modelforge-llmops
7
How do we adapt models to domain format?
RAG + QLoRA triage pipeline (feeds ModelForge)
domainforge-rag-peft
8
How do we run real-time voice triage?
ASR → LLM → TTS with latency budgets
voiceforge-assistant
9
How do we operate agent fleets?
Missions, traces, eval gates
aegisloop-agentops-workbench
11
How do agents improve?
Loop harness · RAG tuning · memory
loop-engine-agent-platform
12
How do we serve LLMs at scale?
PagedAttention · continuous batching · KV budget
vllm-architecture-lab
13
How do we answer multimodal asks with the right models?
Multi-agent multi-LLM routing lab
omniforge
Live Platforms
Click in. Read traces. Evaluate architecture decisions.
Each card = decision + signal + limitation + platform + repo
01Agent governanceGatewayHITLPolicy
AegisAI
Runtime control plane for agent fleets — gateway, policy, HITL, and audit.
Monitor → Govern → Remediate — a runtime control plane for tool authorization, policy violations, HITL approvals, signed audit, and agent registry lifecycle.
Decision
Control plane over agent builder — enterprises need governance across fleets, not another chat UI.
Signal
Website deploy tools forced through approval_required policy on live platform.
Limitation
Registry defaults to in-memory on free tier; Postgres path documented for production.
What shipped: AegisAI: gateway-first governance control plane →
- AI Gateway SDK with OPA policy evaluation
- Human-in-the-loop for high-risk tool calls
- Agent registry with in-memory persistence (Postgres planned)
- Governed orchestrators: content pipeline, stock research
02Multi-agent OSLangGraph16 IntentsRAG Lab
Venkat AI Platform
Multi-agent OS with three orchestrators, seven RAG strategies, and gateway-wrapped delivery.
Three LangGraph orchestrators, seven RAG strategies, loop patterns (ReAct · Reflection · Plan-Execute), and gateway-wrapped delivery to Slack, Telegram, and WhatsApp.
Decision
LangGraph stateful graphs over linear chains — enterprise workflows need checkpoints and HITL.
Signal
16 routed intents · 7 retrieval strategies · AegisAI gateway on notify channels.
Limitation
Vector store optional on free tier; Enterprise RAG adapter bridges to governed RAG platform.
What shipped: Venkat AI Platform: three LangGraph orchestrators live →
- Platform · Deep Research · Architecture Review orchestrators
- Enterprise RAG adapter as 7th retrieval strategy
- AegisAI gateway on notify channels
- Specialist agents: Web, Knowledge, Critic, Planner
03Knowledge layerHybrid RAGAccess ControlGraph Expansion
Enterprise RAG Platform
Access-aware hybrid RAG with ingest, citations, and optional HITL gates.
Authorization before ranking, hybrid retrieval, cross-encoder rerank, decline-to-answer, citation traceability, AegisAI HITL bridge, and Langfuse trace export.
Decision
Authorization before ranking — vector DB is implementation; access control is architecture.
Signal
golden-eval-registry CI · adversarial suite · JWT Principal when PRODUCTION_STRICT=1.
Limitation
Default live demo is Demo mode (body Principal) with sticky banner; Strict/JWT is the Principal review path (ADR-0006). Free-tier corpus re-ingests after cold start.
What shipped: Golden eval registry gates real CI builds →
- Access-aware filtering before semantic ranking
- Hybrid lexical + semantic retrieval with cross-encoder rerank
- Decline-to-answer when retrieval confidence is low
- JWT-verified Principal under PRODUCTION_STRICT · adversarial golden suite
04Model PlaneSLMPEFTvLLM
ModelForge
Model Plane control surface — SLM, PEFT, CUDA vLLM, LLMOps — peer to the agent spine.
Hire-facing Model Plane for SLM bake-offs, PEFT receipts, CUDA vLLM metrics, and LLM gateway enforce+record. Composes DomainForge + upstream vLLM + aegis-llm-gateway (ADR-034).
Decision
One ModelForge flagship beats burying PEFT/vLLM in a teaching drawer (ADR-034).
Signal
Live https://modelforge-gamma.vercel.app/api/v1/posture — PEFT + CUDA vLLM + SLM + gateway all ready; both peft_gpu.json and vllm_cuda.json are real Mistral-7B-Instruct-v0.3 receipts on a rented L4 — real QLoRA SFT + DPO training (peft_gpu.json) and real upstream vLLM serving (vllm_cuda.json, 13.74 tok/s, TTFT p50 371.67ms).
Limitation
PEFT receipt reports real training config/timing, not a quality/win-rate score — DomainForge's S0-S4 eval harness isn't wired to real adapter inference yet (see the receipt's own known_gaps). vLLM metrics are a single dated benchmark run, not an always-on production serve claim.
- Honest /api/v1/posture (ready vs smoke vs planned)
- Receipt gallery for PEFT · CUDA vLLM · SLM bake-off
- Taxonomy glassbox — LoRA · QLoRA · Multi-LoRA · classical ML lane
- Composes DomainForge train + gateway route + vLLM serve path
- Panel scripts for buy vs RAG vs PEFT vs self-host
05Knowledge + MLOpsQLoRARAGEval Harness
DomainForge
Governed support triage — RAG corpus + QLoRA JSON envelope + S0–S2 eval compare.
PEFT training system that feeds ModelForge: SOP RAG, Bitext QLoRA + DPO, S0→S4 eval ladder, Ollama bench UI, adapter registry — RAG holds facts, PEFT holds JSON discipline.
Decision
RAG for facts, LoRA for schema — never memorize policies in weights.
Signal
S0→S4 compare · DPO preference win-rate · `/bench` Ollama metrics UI · `VLLM_BASE_URL` Path B.
Limitation
Mistral QLoRA requires GPU; Render API uses mock/template baseline unless Ollama/vLLM Lab wired. Path B is educational (not CUDA multi-LoRA). Hire hero for models is ModelForge.
What shipped: DomainForge: RAG + QLoRA support triage pipeline →
- 13 SOP docs — hybrid governed retrieval
- Bitext → QLoRA SFT + DPO alignment (S3/S4)
- S0→S4 compare harness + preference pairs
- Ollama bench UI · GPU pipeline docs · educational vLLM Path B (ADR-022)
06Voice / MultimodalASRTTSWebSocket
VoiceForge
Closes portfolio Pillar 5 — measurable ASR / LLM / TTS latency with honest free-tier defaults.
Browser-first ASR on Render free tier, pluggable LLM (mock / Ollama / DomainForge), edge-tts with browser fallback, per-phase latency waterfall, and graceful degradation to text.
Decision
Browser ASR on Render; server Whisper opt-in on GPU hosts.
Signal
9-test CI · P50/P95 latency waterfall UI · phase budgets (ASR/LLM/TTS).
Limitation
Browser ASR quality varies; mock LLM default on Render until Ollama host wired.
- Browser Web Speech API ASR with client-reported latency
- Mock / Ollama / DomainForge LLM triage routes
- edge-tts server synthesis + browser speechSynthesis fallback
- REST and WebSocket transport with phase events
07AgentOps workbenchMissionsTracesEval Gates
AegisLoop AgentOps
AgentOps workbench — bounded missions, traces, eval gates, and human-gated ship.
Bounded missions, specialist handoffs, observable traces, source coverage scoring, FinOps estimates, Langfuse spans, and human-gated ship paths.
Decision
Missions over monolithic prompts — specialist handoffs with observable traces.
Signal
mission_gate CI suite · P50/P95 + failure rate on /api/metrics.
Limitation
Postgres persistence optional; mission history resets on in-memory backend.
What shipped: AegisLoop: mission fleets with eval gates →
- Mission orchestrator with specialist agent routing
- P50/P95 latency + failure rate metrics API
- VAP A2A delegation and AegisAI gateway on ship
- Golden eval CI gate via golden-eval-registry
08Content automationLangGraphHITLPublish
AI Content Factory
Governed content pipeline — research, multi-platform drafts, HITL, then publish.
Research → five platform drafts → HITL review → governed publish through AegisAI gateway, OAuth adapters, and scheduled cron pipelines.
Decision
AegisAI gateway blocks publish until policy allows — side effects never bypass governance.
Signal
Real LinkedIn/X OAuth publish path when tokens configured.
Limitation
Clerk keys required for full pipeline; invite-only signup in production config.
What shipped: AI Content Factory: governed multi-platform publish →
- End-to-end research and multi-platform draft generation
- Clerk auth with human approval gates before publish
- LinkedIn and X OAuth when tokens configured
- AegisAI gateway blocks publish until policy allows
09Self-improvementLoop EngineeringLangGraphMCP
LoopForge
Self-improving agents — clone repos, run pytest, patch, and open PRs.
Agent → Harness → Loops → Memory. LangGraph coding loop (Orchestrator · Review · Quality), ODAEU RAG tuning, and repo fix workflow: clone → pytest → patch → branch → PR — never pushes to main.
Decision
Harness + eval loops over bigger prompts — ODAEU tunes RAG on failure.
Signal
repo_fix suite in golden-eval-registry · never pushes to main.
Limitation
Requires GitHub token for repo fix loop; Groq API for agent inference.
What shipped: LoopForge: clone → pytest → patch → PR workflow →
- Repo fix loop: clone GitHub repo, run pytest, apply patches, open PR
- LangGraph agents: Orchestrator, Coding, Review, Quality, Memory, Self-improve
- ODAEU harness tunes hybrid RAG on eval failure + procedural memory
- Live API on Render · demo UI on Vercel · Groq + GitHub token
10LLM inferencePagedAttentionContinuous BatchingKV Cache
vLLM Architecture Lab
Interactive vLLM architecture explorer — PagedAttention, batching, KV budget, adapter chat.
Educational vLLM reference: PagedAttention, continuous batching, KV budgets, multi-LoRA economics — pairs with DomainForge adapter training (ADR-022).
Decision
Educational simulator before GPU spend — teams learn inference architecture without H100 budget.
Signal
Live simulator · KV budget calculator · DomainForge↔vLLM Path B (ADR-022 Accepted).
Limitation
Simulator + educational adapter swap — not CUDA multi-LoRA kernels or production SLAs.
What shipped: vLLM Architecture Lab: PagedAttention explorer →
- PagedAttention block allocator with swap and prefix sharing
- FCFS scheduler — waiting / running / swapped queues
- KV memory budget calculator (H100 + Llama-3)
- Live API on Render · architecture explorer on Vercel
- Educational multi-LoRA Path B — `/v1/chat/completions` + `/v1/adapters`
11Overnight intelligenceLangGraphCronEval Gate
Sentinel Brief
Overnight intelligence brief — nine sources, snapshot diff, eval gate, governed email.
Allowlisted HN, arXiv, and industry RSS → snapshot diff → executive brief → eval gate → AegisAI gateway email — autonomous inner loop, governed outer loop on mail only.
Decision
Autonomous inner loop, governed outer loop — only email send goes through gateway.
Signal
brief_gate eval · durable archives in /reports · nightly-smoke CI.
Limitation
Render cold start may delay first live run; archives prove report UX without a fresh cron.
What shipped: Sentinel Brief: governed overnight intelligence brief →
- Nine allowlisted sources — RSS/API first
- Snapshot deltas reduce noise vs full feeds
- Eval gate before email; gateway on send only
- Committed archives merged into GET /reports · nightly golden smoke
12Multimodal multi-LLMMulti-LLMMulti-AgentMCP
OmniForge
Multimodal ask with proven multi-LLM routing — waterfall shows which model ran each step.
Self-contained multimodal multi-agent multi-LLM platform — text, image, and voice fan out across agents and MCP tools; task-class routing with a live model waterfall and A/B proof.
Decision
Self-contained monorepo — no sibling runtime deps; routing is architecture, not a dropdown.
Signal
Live demo · /v1/ask waterfall · /v1/ask/ab · ADR-027
Limitation
Render cold start on free tier; mock fallback when cascade misses; clean omniforge.vercel.app may need SSO off.
- Multimodal ask — text · image/screenshot · voice transcript
- Planner fan-out across web/api/data/analysis/vision + MCP tools
- Task-class Multi-LLM Brain — Groq · OpenAI · Anthropic · Gemini cascades
- Model waterfall + A/B routed vs single · in-repo FinOps + export gate
13Interview practiceLLM-as-JudgeSystem DesignInterview Prep
AI Architect Practice Arena
Mock system design and behavioral interviews with dual LLM judges.
Sectioned system-design forms, STAR behavioral practice, and reasoning trade-offs — graded by OpenAI and Anthropic against Staff+/Principal rubrics. Bring your own API key; keys never touch our servers.
Decision
Per-question rubrics over generic chat — grading matches playbook depth.
Signal
139/140 live cases passed · judge disagreement shown, not averaged away.
Limitation
Requires user's OpenAI or Anthropic API key; no server-side key storage.
What shipped: Practice Arena: 35/35 interview playbook coverage →
- 35/35 interview playbook coverage across three rubric formats
- Dual-judge grading with disagreement surfaced as trust signal
- 139/140 live calibration cases passed across two independent runs
- BYOK architecture — zero org-side API cost at any scale
Eval discipline
golden-eval-registry — 10 suite kinds with real CI gates
Spine and AgentOps consumers wire critical suites as merge gates: RAG answer + adversarial (Enterprise RAG), harness QA + repo fix (LoopForge), mission gate + collaboration scorecard (AegisLoop), graph HITL (Content Factory), brief gate (Sentinel), triage preference (DomainForge), and router_invariant across VAP, AegisAI, and FinOps. Builds fail on regression — not just fixture validation.
Architecture Decisions
Why each layer exists — not just what it does.
Canonical ADRs and case studies live in the GitHub architecture portfolio— browse trade-offs before forking code. Architecture theses →
Canonical Essay
From Multi-Agent OS to Agent Governance
Why VAP and AegisAI are complementary layers — not competing products.
Read the full architecture storyFrom Multi-Agent OS to Agent Governance
Orchestration without governance is a liability.
VAP answers what agents should do. AegisAI answers what they are allowed to do. The portfolio ships both — wired together through gateway SDK, HITL overrides, and signed audit — because enterprise programs need both layers from day one.
Proof: Live integration: VAP notify · RAG ingest · Content Factory publish · AegisLoop ship
RAG is not retrieval infrastructure
It is an access-controlled intelligence layer.
Enterprise RAG filters by principal clearance and group membership before ranking — then hybrid retrieval, reranking, graph expansion, and citation traceability. Vector DB selection is an implementation detail; authorization-before-ranking is the architecture decision.
Proof: enterprise_rag_platform · enterprise-rag-platform-eta.vercel.app
AgentOps is how fleets survive production
Missions, not monolithic prompts.
AegisLoop models bounded missions with specialist handoffs, source coverage scoring, evaluation gates, FinOps estimates, and human-gated ship — the operating layer between orchestration platforms and production agent fleets.
Proof: aegisloop-agentops-workbench.vercel.app
Curriculum patterns teach the loops — platforms govern them
ReAct · Reflection · Plan-Execute · Multi-Agent · Swarm — each with a real, CI-gated benchmark.
Five MIT-licensed curriculum stubs, deterministic by design (no LLM API keys) — but each now proves its own pattern's actual guarantee with a real, executed benchmark against a real baseline: bounded-vs-unbounded loop safety, draft-vs-revised score delta, plan decomposition accuracy vs a reactive baseline, reviewer-gate pass rate, fan-out convergence. The numbers aren't uniformly flattering — plan-execute's hard-stop design scored lower than its own reactive baseline on one task set, reported as-is, not smoothed over. Not production fleets — they map ideas into VAP / Content Factory governed graphs.
Proof: 5 live trace viewers + 5 real benchmark receipts, CI-gated via golden-eval-registry
Loop engineering is the next layer
Agents that improve after eval failure — not bigger prompts.
LoopForge implements Agent → Harness → Loops → Memory: LangGraph repo fix (clone → pytest → PR), ODAEU RAG tuning, and procedural memory — live on demo-omega-taupe.vercel.app + loopforge-api.onrender.com.
Proof: demo-omega-taupe.vercel.app · loopforge-api.onrender.com
2026 Agent Protocol Stack
Skills → MCP → Gateway → Observability
MCP for tool access, vpeetla-ai-skills for engineering discipline, AegisAI for side effects, trace-linked LLMOps (Langfuse/OTel) for evals on the same trace_id. ADR-007 documents how platform repos compose for production agent deployments.
Proof: vpeetla-ai-skills · ADR-007 · ORG_IMPROVEMENT_PLAN_2026
Free stack, production boundaries
Vercel + Render + Groq/Upstash/Qdrant — architecture discipline without enterprise budget on day one.
Every live demo runs on free-tier infrastructure with explicit service boundaries: UI on Vercel, APIs on Render, LLM via Groq, vectors via Qdrant Cloud, cache via Upstash — the same pattern I publish in Building Multi-Agent Systems on a Free Stack.
Proof: 15 live products on Vercel + Render free tier
10→2
Agent ops staffing reduction (targeted flows)
Multi-$M
Revenue & savings — payments, subscriptions, EDI
14 live
Reference systems with platforms
37 ADRs
Documented architecture decisions
Architecture Decision Records
Real decisions from production systems
ADR-001
Orchestration vs governance split
VAP + AegisAI as complementary layers — orchestration without governance is a liability.
VAPAegisAI
ADR-002
Authorization before ranking
RAG is an access-controlled intelligence layer, not a vector DB wrapper.
Enterprise RAG
ADR-003
Mission-based AgentOps
Bounded missions with eval gates — fleets survive production, not monolithic prompts.
AegisLoop
ADR-004
Gateway + HITL for side effects
Side-effecting tool calls require policy + human approval + signed audit.
AegisAIContent Factory
ADR-005
Reference stack on free tier
Vercel + Render + Groq — production boundaries without enterprise budget on day one.
Full stack
ADR-006
Loop harness over monolithic agent
Self-improvement belongs in harness + eval + memory — not bigger system prompts.
LoopForge
ADR-007
2026 agent protocol stack
Skills, MCP, gateway, and trace-linked observability as one four-layer stack, not ad hoc wiring.
Full stack
ADR-008
Real publish scope and invite-gating
LinkedIn/X get real OAuth publish; platforms with no public API get honest copy-draft export instead.
Content Factory
ADR-009
Auth gate on VAP routes
Chat, orchestrator, ingest, RAG, and thread routes now require a caller identity, not just a claimed one.
VAP
ADR-010
Auth gate on AegisLoop mission-run routes
Closed the same gap independently in both the FastAPI backend and the Netlify serverless function.
AegisLoop
ADR-011
AgentFinOps as a standalone service
Real cost governance needs one shared ledger, not per-repo fragments — built as its own repo, not logic bolted onto two existing ones.
AegisAIAegisLoop
ADR-012
Real FinOps metering wired into both consumers
A shared cost-governance service only matters once it's actually called — real usage now halts real agent dispatch in both AegisAI and AegisLoop.
AegisAIAegisLoop
ADR-013
Bidirectional MCP + real A2A discovery
A gateway that only gates outbound MCP calls is half a protocol — AegisAI now exposes governed tools via MCP too, and AegisLoop is the org's first client to actually discover-then-invoke VAP's real A2A surface.
AegisAIAegisLoopVAP
ADR-014
golden-eval-registry becomes a real CI gate
Fixtures existing and fixtures being correct are different claims — running one suite for the very first time against a real pipeline found and fixed a genuine bug in the fixture itself.
golden-eval-registryEnterprise RAGAegisLoop
ADR-015
Genuine hands-on AWS + GCP infra
Terraform files nobody ever ran aren't infrastructure experience — a real apply, a real budget breach against a real database, a real orchestrator run against a real load balancer, then a real teardown confirmed empty by each cloud's own CLI.
agent-finopsAegisAI
ADR-016
Ingestion data contracts + real lineage
Validation logic that gets computed and then discarded isn't a data contract — enterprise_rag_platform now rejects bad ingests instead of silently indexing them, and found a whole CI test file that had never actually run along the way.
Enterprise RAG
ADR-034
ModelForge as Model Plane flagship
Agents call tools; the model plane decides buy vs RAG vs PEFT vs self-host — one hire-facing UI beats burying CUDA receipts in teaching drawers.
ModelForgeDomainForgeaegis-llm-gateway
Case Studies
Architecture context, trade-offs, and outcomes
Agent governance
AegisAI — Agent Governance
Runtime control plane — gateway, HITL, signed audit
Includes live demo link →
Multi-agent orchestration
Venkat AI Platform
3 LangGraph orchestrators · 7 RAG strategies
Includes live demo link →
Knowledge layer
Enterprise RAG Platform
Authorization before ranking · hybrid retrieval
Includes live demo link →
AgentOps
AegisLoop AgentOps
Mission fleets · eval gates · FinOps · Langfuse
Includes live demo link →
Model Plane
ModelForge — Model Plane
SLM bake-off · CUDA PEFT/vLLM receipts · gateway enforce+record
Includes live demo link →
Content automation
AI Content Factory
Research → drafts → HITL → governed publish
Includes live demo link →
Loop engineering
LoopForge — Self-Improving Harness
LangGraph repo fix → PR · ODAEU RAG tuning · MCP · memory
Includes live demo link →
LLM inference
vLLM Architecture Lab
PagedAttention · continuous batching · KV budget simulator
Includes live demo link →
Architecture essay
From Multi-Agent OS to Agent Governance
Why VAP and AegisAI are complementary layers
Lucid Motors
Enterprise Agentic AI
10→2 staffing in targeted supply-chain flows
Volvo Cars
Gulf Payments Modernization
Multi-$M annualized revenue impact
Volvo Cars
Subscription Revenue Platform
Durable recurring revenue platform
Volvo Cars
Supply Chain EDI Re-Platforming
Multi-$M annualized savings
Agent Patterns
5 patterns · 5 trace viewers
Mapped to VAP orchestrators — inspectable, not abstract.
ReAct
Thought → action → observation loops with bounded iterations and tool contracts.
Used in: VAP Deep Research orchestrator
Reflection
Critique-and-revise before delivery — quality gates for draft-heavy copilots.
Used in: VAP Research + Architecture pipelines
Plan-Execute
Explicit planning phase before tool execution — safer multi-step enterprise workflows.
Used in: VAP Architecture Review orchestrator
Multi-Agent
Orchestrator-controlled collaboration with reviewer gates before final output.
Used in: VAP Platform orchestrator + specialists
Swarm
Parallel specialist exploration with handoff-friendly agent boundaries.
Used in: VAP parallel asyncio bundles
Open Source
17 catalogued repos · one reference stack
Platform Repositories
View all on GitHubmodelforge-llmops
ModelForge — the Model Plane spine layer: composes DomainForge QLoRA/DPO training, upstream CUDA vLLM serving, and an SLM-vs-API bake-off behind one hire-facing posture surface and receipt gallery.
Live demo- Real GPU receipts: vllm_cuda.json (Mistral-7B-Instruct-v0.3 on a rented L4, 13.74 tok/s) + peft_gpu.json (real QLoRA+DPO training)
- SLM bake-off: local Ollama vs cloud API, schema-pass + latency, honest about what's captured vs deferred
vllm-architecture-lab
vLLM architecture lab — PagedAttention simulator, continuous batching scheduler, KV budget API, and interactive 5-tab architecture explorer.
Live demo- BlockSpaceManager + prefix cache (educational PagedAttention)
- LLMEngine step loop with FCFS scheduler and preemption
vpeetla-ai-skills
Org-wide agent skills for Cursor and Codex — MCP, gateway, loop engineering, TDD, honest status tables, vLLM inference.
- 18 composable SKILL.md packages
- install.sh for Cursor + Codex (AGENTS.md, CONTEXT.md)
golden-eval-registry
Versioned golden fixtures for the governed agent stack — portable, reviewable eval contracts so Enterprise RAG, LoopForge, AegisLoop, and Content Factory can prove they don't regress, not just demo.
- 10 suite kinds with real CI gates — including collaboration_scorecard (CSS / TUE / hard gates)
- Dependency-light, provider-agnostic scorer — consumers fail merges on regression, not just fixture validation
ai-architect-practice-arena
An LLM-as-judge mock-interview tool covering all 35 questions in the interview playbook -- sectioned system-design forms, full STAR practice for behavioral questions, and framework-plus-example for reasoning trade-offs -- graded by both OpenAI and Anthropic against each entry's own real Staff+/Principal rubric, bring your own API key, never stored on or sent to our servers.
Live demo- Full 35/35 playbook coverage across three rubric formats -- sectioned system-design (with a live Mermaid diagram input), full STAR behavioral practice, and reasoning-framework trade-offs -- as one persistent sidebar + form + results shell, not three separate tools
- Dual-judge grading (OpenAI + Anthropic) with per-section strengths/improvements feedback; judge disagreement surfaced as a real trust signal, not silently averaged away
loop-engine-agent-platform
LoopForge — self-improving agent harness. LangGraph repo fix → GitHub PR, ODAEU RAG tuning, MCP tools, procedural memory. Agent → Harness → Loops → Memory.
Live demo- Repo fix loop: clone → pytest → patch → branch → open PR
- LangGraph: Orchestrator · Coding · Review · Quality agents
aegisai-enterprise-agent-platform
Enterprise agent governance control plane — gateway SDK, policy violations, HITL approvals, signed audit, agent registry, and governed orchestrators.
Live demo- Monitor → Govern → Remediate control plane UX
- AI Gateway with OPA policy and human-in-the-loop paths
agent-finops
Real, enforced cost governance for AI agent fleets — usage metering, budgets, and breach signals as a shared service, wired into both repos that used to fake it.
- One canonical pricing table + real per-call token metering, not guessed cost
- Live in production: AegisAI's Website Build agents halt via a real kill-switch on breach; AegisLoop's research/content missions halt via a budget guard
venkat-ai-platform
Principal-architect multi-agent OS — 3 LangGraph orchestrators, 7 RAG strategies, loop patterns, and gateway-wrapped delivery channels.
Live demo- Platform · Deep Research · Architecture Review orchestrators
- Enterprise RAG adapter as 7th retrieval strategy
enterprise_rag_platform
Production RAG as a governed intelligence system — access-aware retrieval, context engineering, evaluation, guardrails, and operational decision records.
Live demo- Authorization before ranking, not after generation
- Hybrid retrieval with citation traceability
aegisloop-agentops-workbench
Mission fleets with evaluation gates — bounded missions, specialist handoffs, traces, FinOps, Langfuse spans, and human-gated ship paths.
Live demo- Mission orchestrator with source coverage scoring
- VAP delegation and AegisAI gateway integration
ai-content-factory
Multi-agent content pipeline — research, five platform drafts, HITL review, governed publish (real LinkedIn/X OAuth; other platforms export a copy-ready draft).
Live demo- LangGraph pipeline with human approval before publish
- Real LinkedIn/X OAuth + PKCE, invite-gated signup
Agent Pattern Library (GitHub)
multi-agent-system-pattern
Centralized multi-agent architecture with specialized roles, orchestrator control, shared context, and reviewer gates before final output.
react-agent-pattern
ReAct pattern for tool-using assistants with bounded iterations, schema-light tool contracts, and separation between reasoning and execution.
plan-execute-agent-pattern
Plan-then-execute agent workflow for decomposing complex tasks into reviewable steps before tool execution begins.
reflection-agent-pattern
Self-critique and revision loop for improving agent output quality before responses reach users or downstream systems.
swarm-agent-pattern
Swarm-style agent coordination for parallel exploration, specialist handoffs, and emergent task routing across agent roles.
Enterprise Delivery
Production delivery outcomes — enterprise context
Agentic AIRAGMLOpsGovernance
Enterprise Agentic AI Platform
Lucid Motors · Automotive Manufacturing · 2023–Present
Enterprise teams don't need another demo. They need repeatable operational work with governance, evaluation, and observability built in.
Cut staffing from 10 people to 2 on the repeatable flows this automated.
PaymentsStripeGIBRetail Commerce
Gulf Payments Modernization
Volvo Cars · Automotive & Global Commerce · 2019–2023
Gulf-market transactions needed payment coverage the existing gateways didn't fit — reliability and revenue were both on the line.
Multi-million-dollar annualized business impact, and a platform foundation that held up for regional growth.
SubscriptionsRevenue SystemsAPIsPlatform
Subscription Revenue Platform
Volvo Cars · Automotive Digital Commerce · 2019–2023
Core product lines were stuck on transactional commerce patterns — no recurring revenue path.
Durable recurring revenue growth — and a platform that actually controls its own subscription lifecycle.
Supply ChainEDIModernizationCost Reduction
Supply Chain EDI Re-Platforming
Volvo Cars · Automotive Supply Chain · 2019–2023
SAP and TrueCommerce EDI workflows were license-heavy — recurring cost, integration friction, and barely any engineering ownership.
Multi-million-dollar annualized savings, and supply chain ops that could actually adapt.
The intellectual rhythm behind the code — weekly architecture essays on Substack and this blog.
Architecture portfolio (GitHub)Read architecture writing