Ben Sicat, AI Engineer
I build AI thatruns in production.
Computer vision, private LLM platforms, and the full-stack systems around them.
M-PCAM
Food volume from a single photo.
M-PCAM estimates how much food is on a plate from one smartphone picture, with no special hardware. It is a hybrid: deep learning finds the food, classical camera geometry measures it.
- Mask R-CNN
- ARCore depth
- Pinhole camera model
- Python
Segment every food item
Mask R-CNN separates each item on the plate, so every portion is measured on its own.
Recover depth
ARCore depth data and a pinhole camera model turn pixels back into real-world distances from the lens.
Measure the volume
Geometric analysis integrates the height of each item over its footprint to get a portion size for dietary tracking.
85-95%
volume accuracy from one phone photo
A private LLM for the whole office.
I built and run the company's own AI assistant on the office network, so staff stop pasting sensitive data into public chatbots. It has tool calling, search over our codebases, and an eval harness to prove each change helps.
58 tok/s
decode on a single RTX 5080, with 633 tok/s prefill
10
custom tools, from read-only text-to-SQL to a Confluence reader
1 gateway
for three groups: developers, a legal pilot, and general staff
- Qwen3 27B on Ollama
- Open WebUI
- Entra ID sign-in
- Self-hosted web search
- Code search: tree-sitter, Qdrant, hybrid ranking, reranker
- Blind A/B eval harness
Scan QC
A scan QC gate that knows when it is unsure.
Every dental case arrives as two or three 3D scans. Scan QC scores the case on CPU from geometric checks, auto-accepts only what it is confident about, and queues the rest for a person, worst first, with findings in plain words like "long open boundary".
- Mesh geometry
- scikit-learn
- FastAPI
- Docker
- Shadow-mode logging
Simple geometry beat a deep net
Same cases, same split. Measured in AUC, where 0.5 is a coin flip.
| Approach | AUC |
|---|---|
| Geometric features + gradient boosting | 0.756 |
| PointNet on raw point clouds | 0.64 |
An early model scored 0.87 AUC. It was reading file formats, not scan quality. The numbers here come from held-out labs and unseen batches.
0.83AUC
pooled over 10,000+ cases it never saw. Batches range from 0.75 to 0.83.
0.80AUC
cross-validated across 97 labs
~97%
of auto-accepted cases are clean
Systems I've shipped.
Most of this is proprietary, so there's no public code. Here is what each one does and how it's built.
Excel, retired
ERP and payroll, vetted through beta
FundPro ERP
A custom web app that replaced FundPro's spreadsheets. It pulls data from their sources, runs payroll, ships preset exports, and puts every number on one overview page.
Next.js on Vercel
60%
less manual CAD design time
Dental 3D tooling
Took an existing tooth-generation pipeline further and built the tools around it: a mesh labeler for training data and an ACAD tool for the lab's technicians.
3D deep learning, mesh processing, ACAD
2,942
conversation chunks labeled locally
Customer conversation labeling
Scrubs personal data from customer chats, filters what matters in three gates, then labels theme, emotion and escalation with local Llama models.
Llama 3.1 8B, Llama 3.2 1B, Ollama, Docker
206 tests
behind a production portal API
Customer portal backend
The integration layer between a lab portal and two data sources, with retries, a circuit breaker, fallbacks, caching and lab-level access control.
FastAPI, Azure App Service, JWT, Locust
<200 ms
inference for live recommendations
Customer preference engine
Reads behavioral signals in real time to predict churn and serve personalized product recommendations.
ML microservices, inference APIs, dashboards
Earlier work.
Research, internships and analyst work that got me here.
Where I've worked.
BS Computer Science, Data Science specialization, Adamson University.
Oct 2025 - Present
AI Engineer
Nimbyx Inc. (Evismart)
The office LLM platform, scan QC and 3D tooling for a dental lab, and the customer portal backend.
Jun 2025 - Oct 2025
Software Engineer & AI Researcher
Veda Technologies
Stateless Common Lisp backends and hybrid symbolic-statistical AI systems.
Oct 2024 - May 2025
Business Analyst
Vertiv Holdings
ETL pipelines from Snowflake and Oracle ERP into executive dashboards, plus supplier data cleanup.
Sep 2021 - Dec 2022
Software Developer Intern
Raya Solutions
A React Native app for car servicing with a Firebase backend and real-time sync.
- Python
- TypeScript
- TensorFlow
- Ollama
- Qdrant
- FastAPI
- Three.js
- Common Lisp
- React Native
- Qwen3
- Llama
- LangChain
- scikit-learn
- Docker
- PostgreSQL
- Redis
- Next.js
- Azure
Building something with AI?
Tell me what you're working on.
bensicat00@gmail.com