Firmatek — Computer Vision for Utility Inspection | Punch
← ALL WORK / AI PROJECTS / FIRMATEK
Firmatek

Every pole, inspected from the sky.

How Punch trained and deployed a computer vision system that grades utility-pole condition from drone imagery — and routes repairs to crews in the field.

COMPUTER VISION AWS SAGEMAKER YOLOV11 DRONE IMAGERY
Firmatek
YOLOV11-S · MAP@50 0.84 · RECALL 0.95 LINE PATROL · DETECTION RUNNING
FROM PIXEL TO LABEL

Labels, versioned like code.

Drone imagery lands in S3 and goes through our in-house annotation pipeline — poles, lines, and insulators boxed by hand, defects graded against a shared taxonomy, every release QA'd before a model sees it. This is the software the labeling team works in:

The annotation suite — poles, lines and insulators boxed by hand
FIG. 00 — THE ANNOTATION SUITE · POLES, LINES, INSULATORS — BOXED BY HAND
The defect taxonomy module — sag, damage, and insulation failures tracked per pole ID
FIG. 01 — THE DEFECT TAXONOMY · SAG, DAMAGE, INSULATION FAILURE — PER POLE ID
WHAT THE MODEL SEES
FIG. 05 — LINE FAULT, SAG & DAMAGE · FLAGGED AT 85 M
FIG. 05 — LINE FAULT, SAG & DAMAGE · FLAGGED AT 85 M
FIG. 06 — INSULATION FAILURE · CAUGHT MID-PATROL
FIG. 06 — INSULATION FAILURE · CAUGHT MID-PATROL
FIG. 07 — EVERY POLE GETS AN ID · PL-472-A
FIG. 07 — EVERY POLE GETS AN ID · PL-472-A
FIG. 08 — POWER LINE FAULT · CRITICAL → DISPATCHED
FIG. 08 — POWER LINE FAULT · CRITICAL → DISPATCHED
FIG. 09 — LINE DAMAGE · PRIORITY HIGH
FIG. 09 — LINE DAMAGE · PRIORITY HIGH
HOW IT LEARNED TO SEE

Only gated models ship.

01 TRAIN SAGEMAKER · GPU
→
02 STORE ARTIFACTS → S3
→
03 EVALUATE HELD-OUT SET
→
04 GATE MAP@50 ≥ 0.78 · MAP@75 · RECALL
→
05 REGISTER → MODEL REGISTRY ✓
↻ GATE FAIL → RETUNE + RETRAIN
FIG. 01 — THE FIVE-STEP GATED SAGEMAKER PIPELINE · RUNS ON EVERY CANDIDATE MODEL
The Roboflow annotation workspace — bounding boxes, two classes, the labeling queue ANNOTATE

Drone imagery lands in S3 and is annotated in Roboflow — bounding boxes, two classes: good / bad.

Seven months of runs — RCNN plateaus at 0.71, YOLOv11 climbs through the 0.78 gate to 0.84 COMPETE

Two architectures trained in parallel — RCNN vs YOLOv11, each with its own pipeline. Months of iterations and tuning. YOLOv11 won decisively.

One batch-transform inference — class, confidence, box, routed straight to the field crew DEPLOY

Lambda triggers batch transform on new imagery. Every inference returns class, confidence, and box — flagged poles route straight to field crews.

THE UNGLAMOROUS MIDDLE · SEVEN MONTHS OF RUNS MAP@50 · SAMPLE VALUES
RUN 0041 · MO 1 0.61 First full pipeline pass. Gate: failed.
RUN 0102 · MO 2 0.68 Label refresh V12. Shadows still fooling it.
RUN 0233 · MO 4 0.74 Synthetic shadow set lands. RCNN plateaus at 0.71.
RUN 0381 · MO 6 0.79 Hyperparameter sweep. First gate pass.
RUN 0492 · MO 7 0.84 Registered. Deployed. The one in production.
THE EVAL · WE SHOW OUR NUMBERS
YOLOV11-S0.84
RCNN / DETECTRON0.71
SAME DATA · SAME GATES · PARALLEL PIPELINES
RCNN RETIRED AFTER PLATEAU
FIG. 02 — THE BAKE-OFF · MAP@50 AFTER FINAL TUNING (SAMPLE)
PRED GOOD
PRED BAD
ACT GOOD
1,842
96
ACT BAD
121
2,403
FIG. 03 — CONFUSION MATRIX · EVAL RUN 0492 · BAD-CLASS PRECISION 0.96 · RECALL 0.95 (SAMPLE VALUES)
0.0241 — TRAIN --- VAL EPOCH 0 → 118 · YOLOV11 FINAL RUN
FIG. 04 — CONVERGENCE · MONTHS OF ITERATIONS, REFRESHED LABELS, TUNED HYPERPARAMETERS (SAMPLE)
WHAT CHANGED

The truck rolls only when the model says so.

Crews used to review drone footage by hand for weeks. Now every flight is graded automatically, and flagged poles arrive as dispatch-ready cards — class, confidence, location.

Low-confidence detections still route to a human. That's the deal: the model does the looking, people make the close calls.

FLAG #4471 BAD 0.82
Crossarm rot — upper third 38.7223 N · 121.3411 W FLIGHT 0492 · 14:32 UTC
Dispatch Crew 04 → Send to review
THE LINEMAN'S VIEW — DESIGNED UI, COMPOSITED INTO RENDERS
WHAT FIRMATEK COULDN'T SEE

Thousands of poles. Inspected by truck, by ladder, by guess.

Firmatek needed to assess pole condition across enormous territories. Manual inspection is slow, costly, and inconsistent — and a missed bad pole is an outage, or worse. They needed a system accurate enough to dispatch real crews on its word.

Punch built the full loop: drone imagery in, graded poles out — every detection carrying a class, a confidence score, and a location a lineman can drive to.

“Every pole scored by the model before a human looks twice.”
31,000IMAGES ANNOTATED
7OF TRAINING ITERATIONS
0.82+DETECTION CONFIDENCE IN PRODUCTION
WHAT WE BUILT
MODELS✓YOLOv11 detection model✓Vector-space RGB classifier✓OralTox strip-reader model
PIPELINE✓Drone-image labeling floor (31k images)✓SageMaker training & evaluation✓Confidence-gated model registry
PRODUCT SURFACE✓Automated pole inspection✓YOLO eval on AWS SageMaker✓Field-ready detection output
SERVICES PROVEN Computer vision systems ML pipeline management (MLOps) Hyperparameter tuning Data labeling & dataset ops Model evaluation & QA CLIENT Firmatek — telecom & power-line infrastructure CASE STUDYfirmatek case study coverCase study PDF ↓IN THE NEWS PIT & QUARRY · OCT 2021 Firmatek acquires Kespry — drone-based aerial intelligence. ↗ COMMERCIAL UAV NEWS · NOV 2021 Firmatek brings its mining expertise to the Kespry platform. ↗ FIRMATEK.COM · OCT 2021 Firmatek announces the acquisition of Kespry. ↗
Punch / Lab NotesISSUE 14
LATEST BUILD
THIS ISSUEInside the models we shipped this month.What we trained, what surprised us, and the eval that let it ship.
PUNCH · AI RESEARCH NEWSLETTERThe AI we’re building next — in your inbox.Research notes from the lab, the models we’re training, and the new projects we’re shipping. A few times a month — no spam, no fluff.★FREE FOR SUBSCRIBERS · THE AI EVAL PLAYBOOK — SEE INSIDE ›
RECENT ISSUESPer-subscriber avatarsYOLO eval at scaleOffline product recognition· JOINED BY 2,400+ ENGINEERS & FOUNDERS
Need work like this? Talk to the engineers who built it
←PREVIOUS — THE LEAGUETHE LEAGUEEngineered to be selective. NEXT — NUMINNUMINTeaching the chart to remember.→