RDKB · Agentic AI Proposal
01 / 12
Proof of Concept · 2026

From Dashboard to Agent.
Automated Firmware Intelligence for RDKB.

A proof of concept for continuous, AI-driven firmware release monitoring — built in one day on real telemetry.

01 · The Problem Today

Firmware releases are monitored manually.

👁

Hands-on review

Engineers must open the dashboard and run comparisons by hand after every release.

⏱

Hours-late signal

Issues surface hours or days after users are already affected.

📊

Doesn't scale

WiFi regressions need expert interpretation — not scalable across all OPCOs.

"The data is there. The insight isn't automatic."
02 · The Opportunity

Every firmware release is the same analysis problem.

"If it's repeatable, it can be automated — and made smarter with AI."
03 · What We Built

A working agentic POC — built in one day on real data.

The Stack
  • LangChain agent w/ function calling
  • FastAPI backend with streaming
  • Vite + React frontend
  • Real RDKB telemetry (event_rate_hourly)
What it does
  • Detects new firmware cohorts automatically
  • Runs A/B comparison without human input
  • Flags WiFi regressions with red/yellow scoring
  • Generates a plain-English RCA report
04 · The Agent Loop

The LLM orchestrates. The tools do the work.

▸ user.select_model()entry point
↓
▸ detect_cohorts()finds Firmware A baseline + Firmware B new
↓
▸ run_ab_comparison()ranks all events by deviation factor
↓
▸ summarize_new_disappeared()flags new + missing event types
↓
▸ get_wifi_deviations()WiFi-only RCA with categories
↓
▸ llm.write_report()plain-English findings
Safe by design: the LLM decides what to investigate and in what order — but the SQL is pre-written and parameterized. No prompt injection risk.
05 · The Cohort Model

A cohort is just model + firmware version.

Firmware A · BaselineFirmware B · New Release
ModelCGM4331COMCGM4331COM
VersionOFW 5.6.0OFW 5.7.0
Devices1,240 MACs387 MACs
Threshold≥ 20 devices≥ 20 devices
Baseline = most recent prior version for the same model. New cohort triggers automatically at ≥ 20 unique MACs. Zero per-release configuration.
06 · Sample Output

Raw telemetry → actionable intelligence in under 60 seconds.

Firmware Release: OFW 5.7.0 on CGM4331COM
Compared against OFW 5.6.0 · 387 devices monitored
WIFI_AUTH_FAIL — 3.2× higher rate. Value avg decreased simultaneously → possible driver issue.
WIFI_DISCONNECT — 2.1× higher rate. No value change. Investigate connection stability.
WIFI_BAND_STEER_FAIL — New event type in OFW 5.7.0. Not present in baseline.
WIFI_WPS_TIMEOUT — Disappeared in new firmware.
▸ Recommendation: hold wider rollout pending WiFi driver investigation.
07 · Full Vision

Four agents. One continuous loop.

🛰

1 · Release Watcher

Polls hourly. Detects new cohorts (≥ 20 MACs). Registers baseline automatically.

⚙

2 · Auto A/B

Runs full comparison the moment a new cohort is detected. Ranks all KPI deviations.

📡

3 · WiFi RCA

Filters to WiFi events. Categorizes driver / hardware / new / disappeared. Structured findings.

📣

4 · Campaign Agent

Routes alerts to teams or OPCOs. Configurable per client. Phone, email, Slack.

08 · POC → Production

The POC is one config change away from live data.

POC · TodayProduction
XLSX export from BigQueryLive BigQuery connection
One-shot on demandHourly scheduled loop
Console + HTML reportLooker tile + Slack alert
Single modelAll models, all OPCOs
Manual UI triggerFully autonomous
Agent logic, tools, and LLM integration are identical. Only the data source and scheduler change.
09 · Living Dashboard

Agents write back to Looker — no new UI needed.

▸ Agent runs comparisonscheduled / triggered
↓
▸ Writes to BigQuerycohort_comparison_results
↓
▸ Looker reads tableexisting connection
↓
▸ "Latest Auto A/B" tileappears in current dashboard
The Looker dashboard becomes the agent's output surface. Zero new tooling for end users.
10 · Scalability

Built for one OPCO. Ready for all.

OPCOStatus
NL · Delin● POC complete — real data
Telenet○ Configuration only
Vodafone (VFZ)○ Configuration only
What changes per OPCO
  • Cohort thresholds
  • Alert recipients
  • WiFi event taxonomy (if different)
What stays the same
  • All agent logic
  • All tools and SQL
  • LLM and framework
11 · Next Steps

Three actions to move from POC to pilot.

STEP 01 · WEEK 1–2

Validate

  • Run POC on 30 days of real Delin data
  • Validate factor thresholds with engineers
  • Confirm WiFi event taxonomy
STEP 02 · WEEK 3–4

Automate

  • Connect to live BigQuery
  • Add Cloud Scheduler · hourly
  • Wire Agent 1 → 2 → 3 pipeline
STEP 03 · MONTH 2

Pilot

  • Deploy for Delin (NL)
  • Add Looker "Latest Auto A/B" tile
  • Run alongside manual process
"The hardest part is done. The logic works, the data is real, and the agent runs."