Conversation Intelligence at Scale
TELUS Digital
Problem. A contact center generated more than half a million real
customer conversations a month. Nobody could read them, so the signal inside them
was lost.
Approach. LLM extraction across every conversation, RAG retrieval
over Qdrant and Turbopuffer, and deterministic heuristics wherever determinism
mattered more than fluency. Parallel execution pipelines with conversation-memory
compression kept per-conversation cost viable at that volume.
- Python
- GCP
- BigQuery
- Vertex AI
- Qdrant
- Turbopuffer
- LangSmith
3x sales conversion · 2x customer retention · 500K+ conversations processed monthly
Agentic Analytics Layer
TELUS Digital
Problem. Analysts needed answers the platform could not anticipate,
at a volume no analyst could reach by hand.
Approach. An agentic layer with intent-based routing, sequential
tool calling, automatic retry and human-in-the-loop validation on critical outputs.
Model quality holds across releases through an evaluation harness built on a
100K+ annotation reference set, human review and production drift monitoring.
- Vertex AI
- Gemini
- LangSmith
- BigQuery
- Python
50 daily users · analyses previously impossible at that volume · 100K+ annotation reference set
Livestock Disease Detection
TELUS Digital
Problem. Field diagnosis depended on manual inspection — slow,
costly, inconsistent, and performed on imagery nobody controlled.
Approach. A detection-then-classification pipeline, RT-DETR followed
by a Vision Transformer, trained for uncontrolled field conditions. A purpose-built
annotation platform took engineers out of the labeling loop entirely.
- PyTorch
- ViT
- RT-DETR
- React
- Node.js
- PostgreSQL
- Docker
1M+ animals classified across 5 categories · 100K+ per month · 3x lower diagnostic cost and turnaround · 12 annotators, 80,000+ images
Offshore Safety Vision
Deloitte · Petrobras offshore safety program
Problem. 20+ offshore oil platforms producing 24+ hours of video a
day on low-resolution cameras over limited bandwidth — conditions where purely
neural approaches degrade.
Approach. Detection, classification and segmentation models for
helicopter and personnel detection, aircraft and worker-role classification, and
hazardous-zone segmentation. Where the networks alone failed on degraded video,
accuracy was held by combining them with contour analysis, object-geometry
heuristics and multi-object tracking.
- Python
- PyTorch
- YOLO
- RT-DETR
- Mask R-CNN
- UNet
- ByteTrack
- OpenCV
- AWS
- MLOps
- CI/CD
Above 80% precision sustained across the fleet · manual video monitoring replaced · incident detection coverage widened
Golf Swing Biomechanics Platform
Independent consulting engagement
Problem. Amateur golfers get feedback that is subjective, delayed,
and impossible to compare against a professional reference.
Approach. A GPU pipeline estimates pose across 33 landmarks per
frame, classifies the eight swing phases with a fine-tuned Vision Transformer,
detects club and ball with RT-DETR, derives 40+ biomechanical metrics per phase,
renders an annotated video synced side by side against a professional reference
swing, and grounds an LLM coaching chat in that swing's own metrics through RAG.
- PyTorch
- MediaPipe
- ViT
- RT-DETR
- FastAPI
- LangChain
- Qdrant
- AWS Lambda
- SQS
- RunPod GPU
- Neon
- Docker
25–40s end to end per swing, including rendering and upload · 2–4s coaching responses · FP16 quantization: 1.39x faster inference at half the model size, confidence preserved to four decimals
Flux Wallet
Personal product · fluxwallet.com.br
Problem. Finance assistants read your data and stop there. The
useful version writes to it — which is also the version that can do real damage.
Approach. A LangGraph agent with 52 tools that call the same
service functions the REST API calls, so an agent-created record is
indistinguishable from a form-created one. Destructive operations are two-turn: a
dry-run summary with row counts and period span, then explicit confirmation. Intent
maps conservatively — "I canceled Netflix" pauses future occurrences rather than
deleting history. Long-term memory runs on Mem0 over Qdrant, and complexity-based
routing sends cheap questions to cheap models.
- FastAPI
- LangGraph
- LangChain
- Mem0
- Qdrant
- PostgreSQL
- React
- TypeScript
- Capacitor
- Cloud Run
- Neon
- Vercel
Production SaaS · per-column AES-256-GCM encryption · enforced multi-tenant isolation · 117 backend test suites and 29 end-to-end specs
Personalized Running Coach Agent
A Brazilian fitness-tech client
Problem. A 500+ member running community needed training that
adapts to level, phase and question. No static plan does that.
Approach. Five or more specialized agents adapt to runner level,
run phase and question type, monitoring smartwatch telemetry live and generating
post-run reports on metrics, progression and injury prevention. Each user's
questionnaire, prompt and generated plan persist as embeddings, so every new plan
is generated with that history as context.
- FastAPI
- LangChain
- ChromaDB
- OpenAI
- Docker
- SOLID
- Dependency Injection
500+ member community served · 5+ specialized agents · live smartwatch telemetry
Retail Loss Prevention Vision
An unattended retail chain
Problem. Unattended stores lose inventory with no way to reconcile
the loss against what actually left the shelf.
Approach. Surveillance footage is ingested off a network share into
a message-queue pipeline, then processed by keypoint-based person detection and pose
estimation, fine-tuned YOLO product detection, BoT-SORT tracking with
re-identification, and video action recognition — associating products to
individuals over time through hand-product overlap, temporal consistency and action
confidence. Burned-in camera timestamps are recovered by OCR with a vision-model
fallback.
- PyTorch
- Detectron2
- YOLO11
- BoT-SORT
- SlowFast
- OpenCV
- RabbitMQ
- RunPod GPU
- Docker
200+ stores · ~20 FPS per stream · 200,000+ annotated product dataset · 70% detection accuracy · in-store monitoring headcount cut
