TrafficMind AI — System Flow Architecture

01 — MASTER DATA FLOW

End-to-End System Architecture

Five sequential tiers — from raw ASTRAM CSV through ML inference, LLM intelligence, REST API, to the React command dashboard. Each tier communicates through typed interfaces.

IEEE Technical Specification

Hover or click any architecture component in the diagram to inspect mathematical models, data schemas, code references, and exact metrics.

02 — MACHINE LEARNING

ML Models — Exact Performance Metrics

Three Random Forest models trained on 80/20 stratified splits from 8,173 ASTRAM incidents. One TF-IDF vectorizer. All serialized to .joblib files, loaded at FastAPI startup for <50ms inference.

Priority Classifier

Random Forest · n_estimators=50 · max_depth=12 · Predicts High / Low incident priority before dispatch

Accuracy88%

Class: Low (1,180 samples)

Precision0.89

Recall0.94

F10.91

Class: High (454 samples)

Precision0.82

Recall0.72

F10.77

Top features: Junction + longitude (~52% combined importance)

RF Classifier Joblib serialized

Road Closure Predictor

Random Forest · Same feature set · Predicts road closure requirement for pre-activating diversion routes

Accuracy93%

Class: No Closure (1,367 samples)

Precision0.93

Recall0.98

F10.95

Class: Closure (268 samples)

Precision0.91

Recall0.72

F10.80

Use: Output fed as input feature into Duration Regressor

RF Classifier Chained Model

Duration Regressor

RF Regressor · Predicts incident duration in minutes · closure prediction used as additional input

MAE~38 minutes

R² Score0.41

OutputContinuous (min)

Targetcomputed_duration_min

Duration cap: 1,440 min (24h) to remove anomalies. Fallback chain: closed_dt → resolved_dt → end_dt → modified_dt

RF Regressor Chained Input

TF-IDF Cosine Similarity

10,000 feature sparse matrix on composite text: {cause} {description} at {address} in {zone} near {junction}. Cosine similarity against all 8,173 events. Returns top-K with historical action taken.

Use in Copilot: New incident query → retrieves closest match → injects that match's response record into Naxerion as evidence. Prevents hallucination by grounding in real outcomes.

max_features=10000 Joblib vectorizer Cosine sim

OR-Tools MIP Resource Optimizer

SCIP Mixed-Integer Program. Objective: Min 100x + 20y (x=officers, y=barricades) subject to reduction target constraint.

Officer reduction rate8.0 min × priority_factor

Barricade reduction rate1.5 min × priority_factor

High priority factor1.5×

Sample (120min, High)3 officers, 16 barricades

Expected reduction72% → 34 min remaining

Solver latency<200ms · fallback grid-search

03 — OFFLINE AI

Naxerion LLM — Fine-Tuned, Offline, Our Own

Not using GPT-4 or Gemini — we built and fine-tuned our own traffic-domain language model. Runs on CPU, zero cloud, zero API keys. Full voice pipeline under 3 seconds.

FLAGSHIP — OUR OWN MODEL

MODEL SPECS

Base: Naxerion AI - MENTOR 7B(our Finetuned Model Under Publication)

Fine-tune corpus: Traffic SOPs + BLR incident history

Quantization: GGUF Q4_K_M (4-bit)

Runtime: llama.cpp — CPU only

RAM: <6 GB

Disk: ~4.2 GB

Latency: <9 seconds per response

Internet: None required

CONTEXT INJECTION

① System Prompt

Traffic ops expert, BLR jurisdiction, 120 officers, 300 barricades available

② TF-IDF RAG

Top-3 cosine-matched incidents with historical_action field

③ ActiveEvent JSON

Current: cause, zone, lat/lon, priority, officers deployed, barricades

OUTPUTS

T-120→T+0 Deployment Playbook

Officer + Barricade Count Advice

Secondary Incident Risk Analysis

Diversion Route Suggestions

Historical Match + What Worked

Recovery Phase Plan

VOICE ROUND-TRIP — FULL OFFLINE — <3 SECOND END-TO-END

①

Speak

Officer mic

②

Whisper STT

tiny.en · <300ms offline

③

Intent Parse

deploy/predict/route/risk

④

TF-IDF RAG

Top-3 historical match

⑤

Naxerion LLM

<2s CPU inference

⑥

Coqui TTS

<500ms audio

⑦

UI + Map

Chat render + polylines

04 — EDGE & CONTEXT

Edge Deployment & Context-Aware State

100% offline capable on a single device. Global publish-subscribe store synchronizes incident context across all 6 screens in <16ms.

Edge Deployment Checklist

All 12 API endpoints run offline

All 5 joblib ML models loaded at startup

Naxerion LLM inference on CPU only

Whisper STT offline (tiny.en)

Coqui TTS offline synthesis

OR-Tools MIP solver local

ESRI ArcGIS tile cache (confirmed working, no key)

start.bat one-click launch — zero setup

OpenAI / Gemini API — NOT NEEDED

GPU — NOT NEEDED (CPU only)

Internet connection — NOT NEEDED

Context-Aware Global State

Pub/sub store.ts — no Redux. setActiveEvent(patch) → all 6 screens re-render in <16ms via useActiveEvent() hook.

① Command Center

② Live Map

③ Forecasting

ActiveEvent

pub/sub store

<16ms sync

④ Ops Planner

⑤ Digital Twin

⑥ AI Copilot ★

Fields:

eventCause · zone · junction · lat/lon · priority · roadClosure · durationMin · officers · barricades · diversions[]

Edge Hardware Requirements

<6GB

RAM Required

~4.2GB

Disk (models)

CPU

No GPU needed

RPi 5

/ NUC / Laptop

05 — REST API

12 API Endpoints

FastAPI 0.137 · Uvicorn 0.49 · Pydantic v2 · Auto OpenAPI at http://127.0.0.1:8000/docs

METHOD	ENDPOINT	MODULE	PURPOSE	KEY RESPONSE FIELDS
GET	/health	System	Backend health + loaded model list	status, data_loaded, models_loaded[]
GET	/api/analysis/summary	Analysis	Aggregate KPI metrics from 8,173 incidents	total_events, avg_duration_min, road_closures_required
GET	/api/analysis/cause-distribution	Analysis	Incident count per cause (11 categories)	[ {cause, count} ]
GET	/api/analysis/hotspots	Analysis	1,000 geo-tagged incidents for Leaflet map	[ {lat, lon, cause, priority, road_closure, address} ]
GET	/api/analysis/timeline	Analysis	Monthly planned vs unplanned trend	[ {month, planned, unplanned, total} ]
GET	/api/analysis/junction-ranking	Analysis	Top 15 junctions ranked by incident count	[ {junction, count} ]
GET	/api/analysis/lessons-learned	Analysis	High-risk junctions + zone + delay causes	top_repeat_junctions[], top_zones[], causes_by_delay[]
POST	/api/predict/predict-all	Predict	RF inference chain → priority + closure + duration	predicted_priority, priority_confidence, predicted_duration_minutes
GET	/api/predict/feature-importances	Predict	RF feature importance for explainability chart	{ "junction": 0.31, "longitude": 0.21, ... }
POST	/api/optimize	Optimize	OR-Tools MIP → optimal officers + barricades	recommended_officers, barricades, expected_reduction_percentage
POST	/api/similarity	RAG	TF-IDF cosine → top-K historical matches	matches[]: { similarity_score, historical_action, address }
POST	/api/simulate	Simulate	Queue-theoretic twin — before vs after intervention	simulation_timeline[], metrics.overall_delay_reduction_percentage

06 — DATA INSIGHTS

Lessons Learned from 8,173 ASTRAM Incidents

Key findings extracted directly from the Bengaluru incident dataset — used to drive Naxerion prompting, OR-Tools constraints, and risk gauge calibration.

Top 5 High-Risk Junctions

SilkBoard Junction flagged for permanent officer deployment — highest avg duration + most road closures.

JUNCTION	INCIDENTS	AVG DURATION	CLOSURES
SilkBoardJunc	54	~112 min	18
MekhriCircle	64	~95 min	12
KRCircle	43	~91 min	11
AyyappaTempleJunc	58	~88 min	9
BTMLayout2ndStage	47	~78 min	6

Cause Analysis

Vehicle Breakdown dominates frequency. Water Logging causes longest delays. Accidents drive road closures.

CAUSE	COUNT / SHARE	AVG DELAY	CAPACITY REDUCTION
Vehicle Breakdown	4,896 (59.9%)	—	25%
Water Logging	—	~145 min	65%
Protest	—	~132 min	70%
VIP Movement	—	—	60%
Accident	—	—	50% + closures
Procession	—	—	55%

Temporal Patterns

Peak hour incidents last40% longer

Highest incident density dayFriday

Monsoon water-logging uplift2.3× higher

Peak hours (morning)07:00–10:00

Peak hours (evening)17:00–20:00

Top 2 features (RF)Junction + Longitude (~52%)

Digital Twin — Sample Result

120-min accident · 6 officers · 25 barricades

Peak queue (no intervention)512 vehicles

Peak queue (with intervention)187 vehicles

Total delay saved8.4 hours

Overall delay reduction67%

ModelM/D/1 queue theory · 5-min steps

Dispatch lagT+10 min

07 — INNOVATION HIGHLIGHTS

Why This Gets Shortlisted

Six precise differentiators — each verifiable in the codebase and backed by real metrics from the ASTRAM dataset.

1

Own LLM

Naxerion fine-tuned · no GPT/Gemini · offline CPU

2

MIP Optimizer

Provably optimal · hard constraints · <200ms

3

Voice AI

STT → Naxerion → TTS · <3s offline

4

Digital Twin

Canvas animation · 67% delay reduction demo

5

Context State

6-screen sync · <16ms · pub/sub · no Redux

6

Zero Cloud

start.bat → full platform · no .env · no keys

Proven by Numbers

✓ 8,173 real ASTRAM incidents — not synthetic data

✓ 88% / 93% classification accuracy — verified by held-out test set

✓ 67% delay reduction — computed from M/D/1 queue model

✓ 72% reduction by OR-Tools MIP — for 120min High priority incident

✓ Junction + Longitude = 52% RF importance — verifiable via /feature-importances

✓ SilkBoard = highest risk junction — 18 closures, 112min avg — data-backed

Demo in 60 Seconds

double-click start.bat

Backend (:8000) + Frontend (:5173) launch

→ Select SilkBoard Junction incident on Command Center

→ AI Copilot: "Generate deployment playbook"

→ Naxerion responds in <2s · TTS reads it aloud

→ Ops Planner: OR-Tools returns 3 officers, 16 barricades

✓ Digital Twin shows 67% delay reduction — offline, zero cloud

TrafficMind AICommand Center