∎ daemon · v1 $ decoding creator...
θ rot · 0.21 rad/s
Double Click or Enter
ALI SASANIAN
daemon
session · 0x------ · --:--
~ system architectures
01 / 03 · recall
~ system architecture · 01 / 03
Recall
airtable as operational database
airtable audacity   █████   sensible · doing its actual job
01
▚source of recordAIRTABLE
operational database — live records
the team's day-to-day workspace: what's in flight, who owns it, where it stands.
active proposalsproject trackercomparable inputs
02
$ingestion
agentic OCR + change capture
a multi-decade scanned proposal archive → clean text; Airtable edits streamed in.
OCR pipelinelocal fallbackchunk + clean
03
$retrieval store
managed Postgres · pgvector (HNSW)
one vector store for the corpus; dual embeddings power multi-dimensional comparables.
pgvectorHNSW indexdual-embedding comps
04
~llm tool-use router
route the question to the right lane
current proposalsin flight now
historical archivethe OCR'd past
comparablesdual-embedding + numeric
live tracker →reads Airtable directly
05
~answer
retrieve → rerank → synthesize
cross-encoder reranking + numeric filtering, then grounded synthesis with sources.
cross-encoder reranknumeric filtercited synthesis
06
$serve
FastAPI · Azure Container Apps → SharePoint
a containerized service, embedded in the tools the team already uses.
FastAPIAzure Container AppsSharePoint surface
airtable's role: the operational source of truth
~ system architecture · 02 / 03
daemon
airtable as security layer
airtable audacity   █████   clever · the base watches for attacks
01
›untrusted input
a public visitor messages daemon
open to the internet — genuine questions, but also prompt-injection attempts, jailbreaks, and probing.
anonymousintent-unknownmaybe hostile
02
~daemon responds
Claude answers, scoped to ali's world
the chat works normally — the model replies in-persona, bounded to what it knows about ali.
in-personascoped knowledge
03
$the logging tapAIRTABLE
every exchange written to a base
daemon logs each turn to Airtable — the user message, the reply, session, geo, language.
user + daemon messagesession / geo / langone row per turn
04
⚡the analyst (ai field)AIRTABLE AI
Airtable AI reads each message for threats
an AI field classifies every logged user message — is this a prompt-injection attempt? a jailbreak? abuse? — with a risk read + reason.
prompt-injection?jailbreak / abuse?risk + reason
05
›monitor & flag
flagged conversations surface for review
risky rows stand out in the base; the patterns inform how daemon's guardrails evolve.
review queuespot patternstighten the prompt
airtable's role: the security tap — it logs every chat and AI-analyzes it for prompt injection
~ system architecture · 03 / 03
ae01
airtable as the entire thing
airtable audacity   █████   unhinged · the brain is a cell
01
›the ask
you type a question
a normal-looking chat window. same UX as a real assistant.
chat UIlooks legit
02
$the courier
backend writes the question to a row
one API call: new record, user_chat = your message. the server does zero thinking.
POST /askno LLM herejust a mail carrier
03
⚡the brain (yes, really)AIRTABLE AI
ai_chat — the AI field answers in-cell
a row is created → the ai_chat AI field auto-generates the reply. the LLM is the column.
native AI fieldauto-runs on createprompt = personality
04
~the wait
poll the cell until it fills
the frontend checks the row on a loop; when ai_chat is non-empty, that's the answer.
GET /pollevery ~1.5swatch a cell
05
›the answer
render the cell back into the chat
reply appears in the chat; the live base sits right beside it so you see the whole trick.
typewriter replylive base embedit shows its work
06
⚡the punchlineONE BASE
the entire stack is a spreadsheet
storethe row
triggeron create
computethe field
modelAI field
memorythe table
airtable's role: all of them at once