WHAT_IT_DOES ARCHITECTURE ADVICE_ENGINE LIE_MATRIX SLOPE_LIES FOUNDER SPECS [ ACCESS_BETA ]
NEURAL_NETWORK_ACTIVE // V1.0

THE CADDIE
IN YOUR POCKET.

Every golfer has stood over a bad lie and guessed.
GLA ends the guesswork. Point. Analyze. Execute.

Computer vision reads your lie. A Claude-powered caddie engine synthesizes lie class, distance, wind, and your skill level into a single, actionable recommendation. And you get caddie-grade advice — adapted to your skill level, your distance, and the conditions in front of you.

REQUEST_BETA_ACCESS SEE_HOW_IT_WORKS
MODEL MobileNetV3Small
LIE_CLASSES 12 types
INFERENCE Server-side / GCP
LAST_LIE_DETECTED Deep_Rough_L2
CONFIDENCE 82.4%
ADVICE_READY YES █
PLAYER_TIER MID_HANDICAP
DISTANCE_TO_PIN 148 yds
REC_CLUB 6-IRON / PUNCH
COMPETITIVE_ADVANTAGE

Every app measures
distance.
None read the lie.

GPS yardage. Shot tracking. Swing analytics. The golf app market is saturated with tools that measure everything except the shot in front of you.

The lie is the variable every golfer deals with on every shot off the fairway — and every golfer has been solving it with instinct alone. Until now.

GLA layers lie intelligence on top of the inputs other apps already handle: distance, wind, club selection. The result is the complete picture — the one a real caddie gives you at the bag.

And it doesn't matter if you're a 28 handicapper learning how to escape a buried lie, or a scratch player fine-tuning carry distance from a tight fairway divot. GLA calibrates its advice to you.

GLA_VS_THE_FIELD
Other Golf Apps
GLA
GPS distance ✓
✓Distance to pin
Shot tracking ✓
✓Club recommendation
Wind data ✓
✓Wind-adjusted advice
Lie condition — ✗
✓CV lie detection (12 classes)
Lie severity — ✗
✓Severity scoring
Technique guidance — ✗
✓Skill-adaptive advice
AI caddie advice — ✗
✓Claude-powered natural language advice
SYSTEM_ARCHITECTURE

PRECISION ENGINEERING

Three systems working in sequence to turn a difficult lie into a confident, executable plan.

01
◈

COMPUTER_VISION

MobileNetV3Small TFLite model classifies 12 distinct lie types — from tight fairway to plugged bunker — with confidence scoring per detection. Rough-medium was deliberately excluded; GLA detects rough and deep rough only, where visual differentiation is reliable. Trained on a proprietary dataset of real-world turf interaction images.

02
◉

ADVICE_ENGINE

Claude Sonnet synthesizes lie class, confidence tier, GPS distance, wind conditions, and player skill level into a single executable caddie recommendation. Ball position, weight distribution, club selection, and shot shape — all derived from what the turf is actually doing to the ball.

03
◎

AUTONOMOUS_PIPELINE

A dedicated edge device runs the TrainClaw pipeline — autonomous dataset validation, Vertex AI training jobs, and CI/CD deployment to Cloud Run. Every beta image that enters the system is validated, labeled, and fed back into the next training run without manual intervention.

ADVICE_ENGINE // v1.0

WHAT THE CADDIE KNOWS

GLA synthesizes up to six inputs to generate each piece of advice. Lie detection is the core capability no other app can replicate — everything else makes it complete.

INPUT SOURCE REQUIRED
Lie class 12-category CV detection
AI model
● CORE
Lie severity Confidence + visual cues
Derived
● CORE
Ball position & stance Weight distribution per lie type
Derived
● CORE
Distance to pin User entry (yards/feet)
User
● CORE
Player skill tier Beginner / mid / low handicap
User profile
● CORE
Club in hand User selection
User
○ OPT
Wind conditions Weather API integration
API
○ OPT
SKILL_TIER_CALIBRATION
BEGINNER
Step-by-step technique for each lie type. Ball position, grip adjustments, swing path, and what to expect from the shot — explained in plain language.
MID_HANDICAP
Club selection rationale, trajectory adjustments, and risk-reward framing. Enough detail to make a confident decision, not enough to overthink it.
LOW_HANDICAP
Carry distance deltas, spin rate expectations, and precision club distance tuning. Fine-grained data to dial in custom yardages and shot shape decisions.

The lie is the foundation. Every other input — distance, wind, club selection — only becomes useful once you know what the turf is actually doing to the ball. GLA establishes that ground truth first, then builds the recommendation around it.

Each lie class maps to a specific setup prescription: ball position in stance and weight distribution. A neutral fairway lie means ball center, 50/50 weight. A downhill lie shifts to 60% front foot. A buried plug in the bunker goes ball back, 70% front foot. GLA outputs the exact split — not vague guidance.

Lie severity cascades through the entire advice chain. A buried lie in deep rough doesn't just change club selection — it changes swing speed, ball position, landing zone expectation, and risk tolerance. GLA models all of it.

Distance anchors club selection. In scoring range, GLA targets the pin. In layup range, it finds the optimal position for the next shot. Carry expectations adjust based on lie severity before any club recommendation is made.

Skill tier changes the voice, not the intelligence. The same 12-class detection, the same physics model, the same inputs — presented at the depth the player can actually use on the course.

SAMPLE_OUTPUT
LIE_DETECTED Intermediate_Rough
SEVERITY Moderate (68%)
DISTANCE 162 yds
BALL_POSITION Center — neutral
WEIGHT_DIST 50% front / 50% back
CARRY_DELTA −11 yds (flyer risk)
CLUB_REC 6-iron → 7-iron
SKILL_TIER MID_HANDICAP
ADVICE //
Neutral setup — ball center, 50/50 weight. Grass will grab the hosel; close the face slightly at address. Expect a flyer: club down one. Land short of the pin and let it release.
LIE_MATRIX // SETUP_PRESCRIPTIONS

EVERY SURFACE.
DIFFERENT GEOMETRY.

Weight distribution isn't a generic tip — it's dictated by the physics of each surface. GLA outputs the exact prescription for your lie. No guessing where to stand.

FAIRWAY NEUTRAL
50% FRONT
50% BACK
Ball positionCenter of stance
StanceShoulder width
Key noteStandard setup — full swing
HARD PAN FORWARD LEAN
60% FRONT
40% BACK
Ball positionSlightly back of center
Shaft leanForward — hands ahead
Key notePick it clean — no divot
ROUGH FORWARD BIAS
55% FRONT
45% BACK
Ball positionCenter to slightly back
GripFirmer — resist twist
Key noteSteeper angle of attack
DEEP ROUGH HEAVY FORWARD
65% FRONT
35% BACK
Ball positionBack of center
SwingAggressive downswing
Key noteCommit — no deceleration
FAIRWAY BUNKER EVEN — PICK CLEAN
50% FRONT
50% BACK
Ball positionCenter — grip down 1"
FeetDig in lightly for stability
Key noteSweep it — don't dig
GREENSIDE BUNKER FORWARD — DIG
60% FRONT
40% BACK
Ball positionForward of center
FaceOpen — aim left of target
Key noteHit 2" behind ball — splash
PLUGGED BUNKER HEAVY FORWARD
70% FRONT
30% BACK
Ball positionBack of center
FaceSquare or closed
Key noteDrive down hard — no follow-through expectation
PINE NEEDLES BALANCED — CAREFUL
50% FRONT
50% BACK
Ball positionCenter — teed up slightly
FeetNarrow — prevent slipping
Key noteHover the club — no grounding
WEIGHT_KEY Neutral (50/50) Forward bias (55–65%) Heavy forward (65–70%) Balanced / surface caution
UNEVEN_LIES // SLOPE_PRESCRIPTIONS

THE SHOTS THAT
BREAK SCORECARDS.

Uneven lies change six variables simultaneously — weight, grip, shoulder angle, aim line, ball position, and ball flight shape. GLA maps all six for each slope condition. No guessing, no half-measures.

UPHILL LIE
Ball above the slope — easier to execute
WEIGHT DISTRIBUTION
60% BACK
40% FRONT
Back foot absorbs slope — match your spine to the hill
BALL POSITION
Center to slightly forward
GRIP
Normal — full grip pressure
SHOULDER ANGLE
Tilted parallel to slope — back shoulder lower
WHERE TO AIM
Right of target (draw bias)
CLUB ADJUSTMENT
Club up 1–2 — slope adds effective loft
BALL FLIGHT SHAPE
DRAW / HOOK BIAS
Slope closes the face through impact — aim right to compensate
DOWNHILL LIE
Hardest of the four — high mishit rate
WEIGHT DISTRIBUTION
60% FRONT
40% BACK
Chase the slope with the front foot — stay down through impact
BALL POSITION
Back of center — promotes contact
GRIP
Firm — choke down 1" for control
SHOULDER ANGLE
Tilted parallel to slope — front shoulder lower
WHERE TO AIM
Left of target (fade/slice bias)
CLUB ADJUSTMENT
Club down 1–2 — slope delofts the face, adds distance
BALL FLIGHT SHAPE
FADE / CUT BIAS
Slope opens the face — aim left, expect lower, hotter flight
BALL ABOVE FEET
Standing more upright — arc goes flatter
WEIGHT DISTRIBUTION
50% FRONT
50% BACK
Sit into heels — gravity pulls you into the hill
BALL POSITION
Center — stand taller, less knee flex
GRIP
Choke down 1–2" — club is effectively longer
SHOULDER ANGLE
More level than normal — flatter swing plane
WHERE TO AIM
Right of target (draw/hook bias)
SWING KEY
Shorter, controlled swing — resist falling back on heels
BALL FLIGHT SHAPE
DRAW / HOOK BIAS
Flatter arc closes the face — aim right, expect right-to-left flight
BALL BELOW FEET
Most physically demanding — balance critical
WEIGHT DISTRIBUTION
50% FRONT
50% BACK
Press into toes — flex knees more, stay bent through impact
BALL POSITION
Center — bend more from hips, reach down
GRIP
Full grip at top — club is effectively shorter
SHOULDER ANGLE
More tilted — steeper swing plane
WHERE TO AIM
Left of target (fade/slice bias)
SWING KEY
Maintain knee flex — do not straighten up through impact
BALL FLIGHT SHAPE
FADE / CUT BIAS
Steeper arc opens the face — aim left, expect left-to-right flight
QUICK_REF // AIM & SHAPE
Uphill
Aim RIGHT
▶ DRAW
Downhill
Aim LEFT
◀ FADE
Ball above feet
Aim RIGHT
▶ DRAW
Ball below feet
Aim LEFT
◀ FADE
Pattern: slope or ground pushing face open = FADE. Face closing through impact = DRAW. GLA calculates the magnitude based on severity of slope.
Johnny Barnachia — Founder, Golf Lie Analyzer
ID: FOUNDER
Program Manager
Solo Founder // GLA
MISSION_LOG

ENGINEERED.
NOT JUST CODED.

MBA PMP® CFCM LSSBB USN_COAST_GUARD_22YR DEFENSE_SECTOR

"I am a Defense & Aerospace Program Manager by trade. I spend my days bridging technical innovation with federal compliance — executing high-stakes hardware/software programs for the DoD. I built GLA because I wanted the same level of data-driven situational awareness I use at work, applied to the most difficult shots in golf."

The discipline that ships flight simulators on compressed timelines for the U.S. Navy — risk identification, systems thinking, zero-defect delivery — is the same discipline behind every architectural decision in GLA.

This is not a weekend app. It is a seriously engineered product built by someone who has spent 20+ years delivering complex systems under real-world pressure, now applying that rigor to a problem every golfer faces on every round.

TECHNICAL_SPECIFICATIONS

CURRENT BUILD STATE

Live specs as of validation Phase 1. Updated with each milestone release.

VISION_MODEL

ArchitectureMobileNetV3Small
FormatTFLite
Lie classes12
Excluded classRough-medium (low detection reliability)
OutputClass + confidence score
SeverityDerived from confidence + visual cues

INFERENCE_STACK

DeploymentServer-side
InfrastructureGCP Cloud Run
Image handlingRetained for model training
PII attachedNone — images are anonymous
Training useLie classification dataset improvement
User consentExplicit at beta signup
Data sold / sharedNever

ADVICE_ENGINE

Core inputsLie class, severity, distance, skill
Ball positionDerived per lie type
Neutral setupBall center, 50% / 50% weight
Weight distributionLie-specific split (e.g. 60F / 40B)
Optional inputsClub selection, wind
Skill tiers3
Advice generationClaude Sonnet — natural language caddie output
Wind integrationWeather API (live)

PLATFORM_STATUS

PhaseValidation Phase 1
Beta capacity500 users
Supported devicesiOS + Android
Priority accessAll handicap levels
Training data useBeta users improve model
DATA_FLYWHEEL

EVERY IMAGE
MAKES IT SMARTER.

GLA's lie detection model improves continuously as beta users capture images from real courses, real conditions, real lighting. The more lies the model sees, the more accurately it reads the next one.

This is the same data flywheel that powers every world-class computer vision system — real-world image volume beats lab data every time. Beta users aren't just testing GLA. They're training it.

Images are retained anonymously — no personal data is ever attached. You consent explicitly at signup. You can withdraw consent at any time. Your data is never sold or shared.

01
USER CAPTURES LIE
Real course. Real conditions.
↓
02
IMAGE ANALYZED
Server-side inference returns advice
↓
03
IMAGE ADDED TO DATASET
Anonymous. Labeled by lie class.
↓
04
MODEL RETRAINED
12-class accuracy improves → repeat
AUTONOMOUS_INFRASTRUCTURE

THE PIPELINE NEVER SLEEPS.

Most AI apps are static — trained once, deployed, forgotten. GLA runs a fully autonomous retraining pipeline on dedicated edge hardware. Every night, DataClaw validates the training dataset. Every beta image that passes quality checks is labeled and staged. When F1 thresholds are met, TrainClaw submits a new Vertex AI training job, evaluates the result, and deploys the winning model to Cloud Run automatically.

This is the same discipline applied to Defense & Aerospace program delivery — continuous validation, zero-defect criteria, automated gatekeeping — applied to a 12-class golf lie classifier. Beta users don't just test the product. They feed the pipeline.

REQUEST_ACCESS

YOUR CADDIE
IS WAITING.

Validation Phase 1 is active. Beta is limited to 500 users across all skill levels — from first-season beginners to scratch players. Every lie image you take feeds the training dataset, making GLA more accurate for every golfer who comes after you. You're not just using the product. You're building it.

No spam. Access notification only. Unsubscribe anytime.

MODEL_TRAINING

Lie images you capture are retained and used to improve GLA's classification model. More images = more accurate lie detection for every user. You consent to this at signup.

ANONYMOUS_BY_DESIGN

No personal data is attached to training images. Your name, account, location, and device ID are never linked to the image data sent for training.

NEVER_SOLD

GLA does not sell, license, or share your image data with third parties — ever. Training data is used exclusively to improve GLA's own lie detection model.

YOUR_EMAIL

Used solely for beta access notification. No marketing lists, no third-party sharing. Unsubscribe at any time.