arx / the AI architect you can talk to

One question.Any of 39 minds.

Pick GPT, Claude, Gemini, Grok, DeepSeek or Kimi. Whichever you choose, Arx grounds the answer in Prince Singh's real production work with hybrid retrieval, remembers the conversation, and streams it in steady 20 ms frames.

POST /cognitiveServices/streamlive

you

idle

GPT 6.1 Sol
Claude Opus 5.5
Gemini 3.8 Flash
Grok 4.7
DeepSeek V4 Pro
Kimi K2.6
GPT 6.1 SolGPT 6 LunaGPT 6 AstraGPT 5.4 ProGPT 5.3 CodexGPT 5.4 MiniClaude Opus 5.5Claude Opus 5.5 FastClaude Opus 5Claude Opus 4.8Claude Opus 4.7Claude Opus 4.6Claude Opus 4.5Claude Sonnet 4.6Claude Haiku 5.5Claude Opus 4.1Claude Sonnet 4.5Gemini 3.8 FlashGPT 6.1 SolGPT 6 LunaGPT 6 AstraGPT 5.4 ProGPT 5.3 CodexGPT 5.4 MiniClaude Opus 5.5Claude Opus 5.5 FastClaude Opus 5Claude Opus 4.8Claude Opus 4.7Claude Opus 4.6Claude Opus 4.5Claude Sonnet 4.6Claude Haiku 5.5Claude Opus 4.1Claude Sonnet 4.5Gemini 3.8 Flash
Gemini 3.6 FlashGemini 3.1 ProGemini 3Google Gemma 3Grok 4.7Grok 4.3Grok 4.2 ReasoningGrok 4.2 Non-ReasoningDeepSeek V4 ProDeepSeek V4 FlashKimi K2.6Kimi K2.5LLaMA 4 MaverickMistral 3Qwen3MiniMax M2NVIDIA Nemotron NanoGemini 3.6 FlashGemini 3.1 ProGemini 3Google Gemma 3Grok 4.7Grok 4.3Grok 4.2 ReasoningGrok 4.2 Non-ReasoningDeepSeek V4 ProDeepSeek V4 FlashKimi K2.6Kimi K2.5LLaMA 4 MaverickMistral 3Qwen3MiniMax M2NVIDIA Nemotron Nano

01 / pipeline

Every question runs
the same 12 steps.

Whichever model you pick, the question goes through one orchestrator: rate check, memory, hybrid retrieval, a prompt in that provider's own format, then a paced stream. Memory is written back after the answer, off the critical path.

tracePOST /api/v1/cognitiveServices/streammodel=claude-opus-5 provider=azure
  1. guard.rateLimit3 questions per 12 h, checked before any model is called
  2. memory.conversationrolling summary of older turns + the last 4 verbatim
  3. memory.userup to 5 facts you told it before, cosine ≥ 0.2
  4. └rag.understandentity detection; “uska architecture” resolves from history
  5. └rag.denseChromaDB nearest neighbours over ~707 chunks
  6. └rag.bm25exact tokens a vector misses: E11000, Kimi K2.5
  7. └rag.fusereciprocal rank fusion, parent-child dedup
  8. └rag.gradeCRAG: weak evidence broadens to a global pass
  9. prompt.buildchat messages, a Gemini prompt or Bedrock Converse blocks
  10. model.streamreader fills a buffer, writer drains it every 20 ms ± 3
  11. memory.commitfolds old turns into the summary, extracts your facts
  12. chat.savestored with sessionId + turn, so the thread comes back

bars show order and overlap, not measured time · dashed = runs in the background after you have the answer

02 / inside

Not a wrapper.
An orchestrator.

retrieval

Meaning and exact words, fused

query

Which Kimi model runs on Bedrock?

dense · by meaning

model catalogmulti-cloud

BM25 · exact tokens

Kimi K2.5Bedrock

RRF + CRAG

fused, deduped, graded, then handed to the model

A vector search alone misses exact names and error codes, so Arx runs BM25 next to ChromaDB and fuses the two rankings. A grader checks the evidence; if it is weak, the search widens and runs again.

fallback

Never a blank answer

  1. 1Search inside the entity you asked about
  2. 2Widen to a global pass
  3. 3Fall back to the nearest entity cards

Salary data stays out of every answer unless the question asks for it directly.

memory

Two layers of memory

this conversation

a running summary of older turns, plus the last 4 word for word

you

facts you state about yourself, deduped, at most 60

Never facts about Prince, never sensitive data. If memory fails, the chat still answers.

streaming

Paced, not dumped

bufferchunk24680 queued

sessions

Conversations, not one-offs

  • Design a RAG pipeline that won't hallucinateClaude Opus 5 · 4 turns
  • How do you keep latency low?Gemini 3.8 Flash · 2 turns
  • Why you over a generic AI?GPT 6.1 Sol · 1 turn

Every chat is its own session and comes back exactly as you left it. Only a session's first question can be answered from cache; follow-ups always go to the model.

access

Free to try, honest about limits

3questions
every 12 hours
email + OTPGoogleGitHub

Run out and you get a live countdown to your next question, not an error page.

03 / models

39 models. 4 clouds.
One picker.

39live

GPT 6.1 Sol/GPT 6 Luna/GPT 6 Astra/GPT 5.4 Pro/GPT 5.3 Codex/GPT 5.4 Mini/Grok 4.7/Grok 4.3/Grok 4.2 Reasoning/Grok 4.2 Non-Reasoning/DeepSeek V4 Pro/DeepSeek V4 Flash/Kimi K2.6/Kimi K2.5/LLaMA 4 Maverick/Mistral 3

{ model: "gpt-6.1-sol", provider: "azure" }

Under the hood

How Arx answers, end to end.

One question, from your browser to the model cloud and back. Click any step to see what runs there and where it lives in the code.

04 / your turn

Bring the hard
question.

AI architecture, RAG, system design, or anything Prince has shipped. Pick any of 39 models and watch it answer from real work.

Open Arx
Prince Singh

Built by Prince Singh · AI Software Engineer & AI Architect · © 2026

Arx | AI Architect Chat - 39 AI Models by Prince Singh