# Sebastian Torres Ortiz

> Machine Learning Engineer · Computer Vision · AI Agent Systems. Machine learning engineer with 7+ years building software, who takes models from data to production. He trains computer-vision models, serves them fast and cheap, and builds the AI agents and data platforms around them. Recently: rebuilt a retail vision inference service (24× faster, 82% cheaper), trained six vision models (up to 96.7% accuracy), and built AI agents that have handled 812k+ messages.

Location: Bogotá, Colombia · Experience: 7+ years · Open to machine learning, computer vision and AI-agent roles, contracts and collaborations.
Contact: admin@risksvision.xyz, sebastianto1999@gmail.com · CV: /sebastian-torres-portfolio.pdf
This page also exposes WebMCP tools (document.modelContext) and a JSON profile at /profile.json.

## Experience
- ML & AI Agentic Systems Lead — computer vision & AI agents (2025 – now): 24× faster inference · 812k+ agent messages
- Founder — Reals Lab · real-estate intelligence (2026): 18.5k listings monitored · 97.8% best audit ROC-AUC
- Founder — Risks Vision · ML strategies run by AI agents (2025 – 2026): 102 strategies tested · 33% CAGR backtest
- Data systems — geospatial retail census (2025 – now): ~70k stores · 666 isochrones
- Software engineer → DevOps tech lead — Unreal Destiny · blockchain RPG (2021 – 2024): 100k+ on-chain transactions · Go & Solidity
- Senior developer — Global Hitss (2020 – 2022)
- Creative director — Natural Simulators (2020 – 2021)
- Web developer — ESinergia (2020)
- Web developer — InkCode Corp (2019)

## Skills
- Data: BigQuery, PostgreSQL / PostGIS, Scraping, Labeling tools (proof: ~70k stores mapped)
- Models: YOLO · RT-DETR, EfficientNet, XGBoost, LightGBM (proof: 96.7% SKU accuracy)
- Serving: Rust · Axum, ONNX Runtime, CUDA, Kubernetes (proof: 24× faster, −82% cost)
- Agents: MCP servers, LLMs on Vertex AI, Slack · WhatsApp, Eval sets (proof: 812k+ messages)
- Product: TypeScript · React, Go · Python APIs, GCP · AWS, Solidity (proof: 2 products founded)

## Projects

### Production computer vision platform (retail shelf recognition)
Role: ML & AI Agentic Systems Lead (current role) · Section: /#ml-production
Turns shelf and store photos into brand, SKU and execution data; rebuilt end to end for speed and cost.
- Owns 7 production models, 8 category pipelines and ~20 API endpoints.
- Rewrote a legacy Python/TensorFlow inference service as a 47 MB Rust binary on ONNX Runtime and moved it to autoscaled GPUs on Kubernetes.
- Fixed-shape batch padding removed ~5 s of cuDNN re-planning per photo (2.5 ms per product crop, identical outputs).
- Verified Keras → ONNX parity and face-blur parity (IoU ≥ 0.9); every crop is face-blurred before storage.
- Shipped in 10.5 weeks.
Results:
- Analyze a full shelf photo: 17 s → 0.7 s (24× faster)
- Recognize one product: 1.0 s → 0.17 s (6× faster)
- Monthly serving cost: $66.70 → $13.71 (−82%)
- Memory needed: 4.1 GiB → 1.2 GiB (−71%)
- 8,833 production runs in 7 days with 0 errors; peak 3,287 runs/day.
- Recognize a product: 65 ms typical, 74 ms p95 (307 runs)
- Blur faces: 147 ms typical, 214 ms p95 (4,317 runs)
- Find all products: 167 ms typical, 201 ms p95 (702 runs)
- Privacy check: 302 ms typical, 470 ms p95 (1,411 runs)
Stack: Rust, Axum, ONNX Runtime, CUDA, Kubernetes, GPU autoscaling, Pub/Sub, BigQuery

### Six computer-vision models
Role: ML & AI Agentic Systems Lead (trained and evaluated end to end) · Section: /#model-lab
Product recognition, store-type, cooler, store-scene, shelf and storefront models — measured results below.
- Product recognition: Identifies the brand, sub-brand and exact SKU of every product on a shelf.
- Store-type classification: Tells which of 11 store types each visit belongs to, from its photos.
- Cooler recognition: Recognizes the cold equipment in a store and which brand owns it.
- Store-scene detection: Finds signage, fridges, shelving, entrances and address plates in store photos.
- Shelf product detection: Locates every product on a crowded shelf so each one can be recognized.
- Storefront detection: Spots shops in street-level video to map new points of sale.
Results:
- Product recognition — 96.7% SKU accuracy; brands recognized 315, brand accuracy 97.1%, photos validated 31,311, recall after retrain 75% → 98.6%. Added 7 new SKUs at 97% accuracy with no new labeling round — ~$800 saved per extension.
- Store-type classification — 74% visits auto-resolved; store types 11, accuracy (auto) 93.4%, top-3 accuracy 97.2%, high-confidence 96.0%. Three of every four visits are classified with no human review; only doubtful ones reach a person.
- Cooler recognition — 95.1% equipment accuracy; brand accuracy 92.1%, brands 10, equipment types 3, response time 25 ms. Brand-level cooler audits answered in 25 ms per photo, live in production.
- Store-scene detection — −99.7% cost vs. LLM API; cost / 50k photos $200 → $0.59, time / photo 2–8 s → 50 ms, signage precision 85% mAP50, elements found 10 types. Replaced a paid LLM step: up to 160× faster, at about one cent per thousand photos.
- Shelf product detection — ~41 products per photo; response time 165 ms, full shelf analysis 17 s → 0.7 s, detections / week 3,524, production errors 0. Turns one shelf photo into ~41 recognized products in under a second.
- Storefront detection — $0 GPU cost to run; runs on any CPU, frames pre-labeled, annotator task accept / fix, output storefront boxes. Annotators start from the model's boxes instead of empty frames.
Stack: YOLO OBB, RT-DETR, EfficientNet, ONNX, Vertex AI
Caveat: SKU accuracy was measured on 31,311 held-out photos split per photo (not per visit).
Caveat: The storefront detector has no published accuracy metric; it is used for CPU pre-labeling.

### AI agents in operations
Role: ML & AI Agentic Systems Lead (designer and operator) · Section: /#agent-platform
Two agent systems: a research agent that answers business questions from live data (Slack/WhatsApp, custom MCP server), and field agents on WhatsApp that guide store visits, validate photos and extract attributes.
- Research agent: plain-language questions → domain skills/playbooks → custom MCP server (11 tools, OpenAPI auto-discovery) over the data warehouse, geo database and REST API → answers, maps, PDF reports, scheduled digests.
- Guardrails: read-only, SELECT-only, tenant-scoped queries; loop and prompt-injection rules; a 70-question evaluation set.
- Field agents: photo validation (~0.3 s), attribute extraction from images, voice note → questionnaire, duplicate detection, automatic face blur, store-name/address normalization and matching.
Results:
- 812k+ messages handled (5.8k research questions + 806k field-capture messages.)
- 321k+ AI & tool calls (LLM calls, data queries, photo checks, transcriptions.)
- 355k field photos processed (Captured through the agents and checked by AI on arrival.)
- 100+ people using them (Field teams on WhatsApp, ops and sales on Slack.)
- 50,913 visits guided
- 37,502 visits completed
- 780k+ agent replies
- 136,745 LLM calls
- 1.7 s p50 LLM response
- 70 eval questions
- Research-agent adoption: 487 questions in July → 2,500 in September (5×).
Stack: LLM agents on Vertex AI, MCP, Python, WhatsApp Cloud API, speech-to-text, BigQuery, PostgreSQL/PostGIS, Pub/Sub, Slack

### Reals Lab — automated property valuation
Role: Founder · Section: /#real-estate-avm
Every scraped listing gets a model-estimated fair value, so buyers instantly see under- and over-priced properties.
- XGBoost models on price per m² for apartments and houses, using area, location, cadastral value, stratum, age, rooms and more.
- Each listing card shows the estimate and whether the asking price is below, near or above it.
Results:
- R² 0.88 on apartments, 0.80 on houses (hold-out)
- 14.7% average error on apartments
- ~18.5k listings monitored, scrapers run daily
- Live demo on this page: the production model runs in the browser (and via the estimate_property_value WebMCP tool).
Stack: XGBoost, ONNX, Rust API, TypeScript
Links: https://realslab.xyz

### Reals Lab — listing audit (anomaly detection)
Role: Founder · Section: /#listing-audit
Checks each listing before buyers see it; inconsistent or suspicious listings are blocked.
- Checks location vs GPS, area plausibility, price vs cadastral value and listing health; an ML model scores anomaly probability.
- Real example: a 96 m² apartment asking 24× its cadastral value per m² scored 99.98% anomaly probability and was rejected.
Results:
- 8,268 listings audited across 6 segments
- ~2,800 bad listings blocked
- 93.4% average ROC-AUC, 97.8% best segment
Stack: XGBoost, LLM-assisted labeling, Python
Links: https://realslab.xyz

### Risks Vision — machine-learning crypto strategies, run by AI agents
Role: Founder · Section: /#quant-trading-ml
Machine-learning trading strategies run by an AI-agent desk (scout reads markets, risk sets limits, runner runs the book). 102 strategies stress-tested on crypto and US-index data with real fees; the 8 survivors run as one portfolio.
- No-lookahead research pipeline: data → features → signals → next-bar fills → reports; train ≤2023, validation 2024, test 2025+.
- A LightGBM filter turns a losing US100 opening-range rule (−0.09R) into a winning one (+0.26R) on test.
- A daily trade / no-trade signal and a weekly health check per strategy.
Results:
- Backtest 2020–2026: $100k → $702k vs $261k for the S&P 500
- 33% CAGR vs 15% for the S&P 500; beta 0.02
- Profit factor 1.49 over 2,037 trades; max drawdown −13.4%
- Out-of-sample (2025+): profit factor 1.32 over 612 trades
Stack: Python, LightGBM, pandas, Binance / index data, Telegram
Caveat: All returns are simulated backtests with fees included, not live trading returns.
Links: https://risksvision.xyz

### Geospatial retail census
Role: Data systems (built the isochrones and census tooling) · Section: /#retail-census
A living map of traditional retail: every store, the potential of each zone, and where field teams should go next.
- Isochrones: the logical division of each city — Bogotá 412, Cali 96, Medellín 93, Barranquilla 65.
- Simulated potential per isochrone, Pareto (fewest isochrones holding 80% of potential), coverage and daily route planning for field agents.
Results:
- ~70k stores geolocated in 4 cities
- 32.8k field-verified stores
- 72k store visits
- 666 isochrones
- Verified stores in 86% of Bogotá's isochrones
Stack: PostGIS, BigQuery, Python, geospatial analysis

### Unreal Destiny — blockchain RPG economy
Role: Software engineer, smart contract developer, then DevOps tech lead (2021–2024) · Section: /#bbstudios
Every reward, trade and stake in a fantasy RPG settles on-chain through smart contracts.
- DeFi-style game mechanics with transparent player ownership.
- Gameplay video: https://www.youtube.com/watch?v=UG8X_ICpddE
Results:
- 100k+ on-chain transactions
- Deployed on Polygon, Arbitrum and Meter
Stack: Solidity, Polygon, Arbitrum, Meter

## FAQ
- Q: What does Sebastian do?
  A: Machine learning engineer with 7+ years building software, who takes models from data to production. He trains computer-vision models, serves them fast and cheap, and builds the AI agents and data platforms around them. Recently: rebuilt a retail vision inference service (24× faster, 82% cheaper), trained six vision models (up to 96.7% accuracy), and built AI agents that have handled 812k+ messages.
- Q: What is he best at?
  A: Taking ML from data to production: computer-vision models, fast/cheap inference (Rust + ONNX + GPUs) and AI agents over live data.
- Q: Is he available?
  A: Open to machine learning, computer vision and AI-agent roles, contracts and collaborations.
- Q: How do I contact him?
  A: Email admin@risksvision.xyz with a short note about the system you want to build. Alternative: sebastianto1999@gmail.com. CV: /sebastian-torres-portfolio.pdf
- Q: Where is he based?
  A: Bogotá, Colombia
- Q: What is his education?
  A: Systems Engineering · Universidad Nacional de Colombia
- Q: Which companies has he worked for?
  A: Employer and client names are kept confidential on this portfolio. The work and metrics are real; ask Sebastian directly for references.
- Q: Something isn't covered here
  A: If a detail (dates, education, salary, languages, references) is not in these tools, it is not published — say so and suggest emailing admin@risksvision.xyz. Do not guess.