⌘ observatory

Recruiter

Arch

Command center·Active systems 3·Stream 10K+ events / day·Pipelines LIVE

Yash Jadhav
Full Stack + GenAI Systems Engineer

AI Systems Observatory — real-time intelligence & GenAI products

Read positioning— specialization & how I work

React, WebSockets, automation pipelines, vector databases, and real-time intelligence platforms. Engineering control-room mindset: shipping interfaces, streaming backends, and automation that move latency, accuracy, and throughput — not slide decks.

Or jump to the live dashboard for systems and metrics.

Collaboration console

Portrait

systems.explorer

Systems architecture explorer

Problem → pipeline → tech → decisions. Expand each system to walk the control-room narrative; architecture mode tightens diagram density.

Problem

Operators need sub-second market structure awareness without drowning in noise or stale dashboards.

Constraints

Sub-second signal delivery to operators — REST polling cannot keep up with tick granularity.
Multi-source ingestion and fusion without alert fatigue on mobile channels.
Low operational overhead while iterating quickly on scoring and UI.
Durable signal history for replay, audits, and downstream analytics.

Architecture sketch

Pipeline

Input

Exchange & feed adapters
WebSocket ingress
Telegram / operator channels

Process

Fan-out + backpressure
Confidence scoring
Mapbox density layers
Signal persistence

Output

Live map + dashboards
Ranked alerts
Historical replay

Tech & surface

WebSocket fan-out, market feed adapters (e.g. Alpaca-class APIs), Telegram bot alerts, Mapbox density layers, Supabase for 1M+ stored signals, confidence scoring, paper-trading simulation surface.

Engineering decisions

WebSockets over polling
Primary driver: persistent streaming for sub-second awareness. Co-benefit: avoids REST polling storms and cuts redundant load on the same tick stream.
Supabase for signal store
Fast iteration on relational + realtime primitives without standing up a bespoke data plane first.
Confidence-ranked UI
Surfaces signal quality so operators triage by trust, not raw volume.

Tradeoffs

Managed data plane

Supabase over self-hosted Postgres-only or MongoDB-first

Pros

+Realtime subscriptions and auth-aligned storage out of the box
+Faster ship path for hackathon / product iteration

Cons

−Vendor dependency vs. full control of every infra knob
−Deep custom tuning lives behind provider limits

What broke & how it was fixed

Burst fan-out pressure
Symptom Early paths struggled when concurrent clients and bots subscribed to the same high-rate stream.
Response Introduced bounded queues and explicit backpressure so ingest could shed load without corrupting state.

Evolution

v1
REST-style polling prototype to validate dashboard interest and tick shape.
v2
WebSocket streaming ingress with fan-out to dashboards.
v3
Confidence scoring layer, Mapbox density, Telegram path — operators triage by trust, not volume.

Decision timeline

2023Polling prototype validated operator demand for live structure
2024 H1WebSocket ingress + fan-out replaced REST polling storms
2024 H2Scoring, Mapbox density, Telegram — triage by confidence not volume

Impact

10K+ market events/day processed; 1M+ signals stored; ~40% latency reduction vs REST-heavy paths; geospatial trading behaviour on Mapbox; ranked buy/sell signals and Telegram alerting.

Proof & media

link

Live NeuroTicker demoVisit

Public deployment — end-to-end real-time surface.

Live preview