01 / SIGNAL
A prompt enters the system.
Noise becomes order. Language becomes signal.
DESCEND TO ENTER
02 / BRANCH
It thinks in branches.
Paths branch and weigh. Weak branches die.
03 / REACH
It reaches for what it doesn't hold.
Meaning clusters in vector space. Move closer — it responds.
Browser, search, database, API — capabilities light up on call.
04 / SWARM
It acts as a team.
Planner, researcher, critic, verifier — messages in flight.
05 / RETURN
Everything converges.
Signals merge. A single, considered answer.
02 — Modules · Active
05 active// modules online — awaiting signal
MOD_01
Client Intelligence · AI Workspace
The AI operating system for financial advisers.
A proactive workspace for financial advisers — client intelligence, meeting preparation, compliance, and document understanding in one operating layer, every answer grounded in sources.
open kritifin.obliviox.in →MOD_02
On-Device AI · Browser Extension
Rewrites rough prompts into clear ones — on your device.
A free Chrome extension: select text anywhere and a small language model refines it in place, on-device by default — WebGPU with a WASM fallback. No account, no telemetry, no backend.
open refineiq.obliviox.in →MOD_03
Knowledge Retrieval
conceptRetrieval · Reasoning
Reads thousands of sources. Returns cited synthesis.
A long-horizon research agent that reads across thousands of sources, verifies claims adversarially, and returns cited synthesis.
MOD_04
Crawl Intelligence
conceptCrawlers · Data Pipelines
Models search intent. Surfaces ranking moves.
A crawler-fed knowledge system that models search intent and surfaces the moves that actually move rankings.
MOD_05
Agent Orchestration
conceptAgents · Orchestration
Agents plan, execute, verify — in parallel.
A deterministic orchestration layer where specialised agents plan, execute, and verify complex work in parallel.
03 — Stack · Instrument
24 keysThe stack, on one board
Twenty-four tools, one per keycap, each carrying load in the work above. Hover one to read it, or just type — the board answers to a real keyboard, and every key has its own note.
// The same board, as a list.
MODEL
RAG
Grounded answers with the sources attached. No citation, no claim.
MODEL
LangChain
The plumbing between model, retriever and tool, so none of it is bespoke.
MODEL
LangGraph
Agents as state machines. A loop you can draw is a loop you can debug.
MODEL
PyTorch
Where fine-tunes and custom heads get trained before anything is served.
MODEL
TensorFlow
Inherited models, kept running and exported where the serving stack wants them.
MODEL
Qdrant
Vectors with payload filters, so retrieval narrows before it ranks.
SERVICE
Node.js
The runtime under the API surface. Streams first, so tokens arrive as they land.
SERVICE
Python
Where the pipelines, the crawlers and the evaluation harnesses live.
SERVICE
FastAPI
Typed request in, typed response out, and the schema is the documentation.
SERVICE
Express.js
Small services that do one thing. Middleware where middleware belongs.
SERVICE
Playwright
Every claim on this site is asserted in a browser before it ships.
SERVICE
scraping
Politely, on a schedule, with the parse pinned so a layout change fails loudly.
DATA
PostgreSQL
One database until it genuinely is not enough. pgvector counts.
DATA
MongoDB
For documents that were never rows — crawls, transcripts, raw tool output.
DATA
Redis
Cache, rate limit and session state. The cheapest answer is the one already computed.
DATA
RabbitMQ
Slow work leaves the request path. Retries and dead letters are the design, not the patch.
DATA
n8n
The glue nobody should write twice — webhooks, schedules, handoffs between services.
DATA
Grafana
Latency, spend and failure rate on one wall. An unmeasured pipeline is a rumour.
PLATFORM
Docker
The same image on a laptop and in production, or it does not count as tested.
PLATFORM
Kubernetes
Once there are enough services that restarting them by hand is a job.
PLATFORM
AWS
Queues, object storage and GPU when the batch is larger than the budget for idle.
PLATFORM
Azure
Where the enterprise tenancy already is, model endpoints included.
PLATFORM
Google Cloud
Run, Pub/Sub and BigQuery, for the pipelines that were already reading from it.
PLATFORM
Jenkins
The pipeline that refuses the merge. Boring on purpose, and never skipped.