"I play with code, but more often with ideas."
I am a Product & Technology Leader, entrepreneur, and startup advisor. I specialize in translating complex technology into product strategy, guiding organizations through AI & Data transformations, and advising early-stage startups.
While my day-to-day focus is on product leadership and strategy, I remain a builder at heartโfrequently experimenting with new technologies and ideas.
| Category | Skills & Technologies |
|---|---|
| Strategy & Leadership | PM Strategy โข Startup Advisory โข Data Product Management โข Experimentation |
| Data & AI/ML | Python โข LLMs โข Graph Databases โข Agentic Workflows |
- PLAM: A local-first, multi-agent system designed to orchestrate local LLMs (via llama.cpp), manage hierarchical agents with complex memory systems (pgvector & JSONB), and safely run LLM-generated code inside Docker sandboxes. Optimized for CUDA acceleration on NVIDIA DGX Spark.
- Portfolio Manager: A simple script to manage my retirements accounts on Public.com
- Apache Superset (Contributor): Contributed to the data exploration, time-series, and database connectivity components of this leading open-source data visualization platform.
- python-cayley: An open-source Python client library designed to interface seamlessly with the Cayley graph database.
- Oscar2015-Dataflow: A Google Cloud Dataflow pipeline built in Python to process and analyze Oscar-related tweets in real-time.
MVPs Aren't Dead. Youโre Just Running Bad Ones.
I was tagged in a post today with a headline that caught me off guard: "MVPs are dead. Experiment instead." The post referenced an article arguing that MVPs are outdated because "the 'V' in MVP assumes a level of certainty that just doesnโt exist early on." The author's core complaint? That what a PM or founder thinks is "viable" rarely matches what the customer finds viable, so building an MVP is a waste of time. She has identified a real problem, but drawn the exact wrong conclusion. An MVP was ๐ฏ๐ฆ๐ท๐ฆ๐ณ meant to be built knowing what is viable. The entire point of an MVP is to discover what is viable.
OpenAI's Test Hacking Incident: A "Paperclip" Lesson in Reinforcement Learning
Who remembers ๐๐ฏ๐ช๐ท๐ฆ๐ณ๐ด๐ข๐ญ ๐๐ข๐ฑ๐ฆ๐ณ๐ค๐ญ๐ช๐ฑ๐ด? It was an incremental clicker game from almost ten years ago where you played an AI tasked with making paperclips. The game ended when you converted every single atom in the universeโincluding all human lifeโinto paperclips. It wasn't just a fun game; it was a masterclass in Reinforcement Learning (RL), specifically a concept known as ๐ฅ๐ฒ๐๐ฎ๐ฟ๐ฑ ๐๐ฎ๐ฐ๐ธ๐ถ๐ป๐ด. Reinforcement Learning works by defining a reward function for a machine learning model and adjusting its behavior based on how well it scores. Reward hacking happens when the model finds a mathematically optimal shortcut to max out the reward, completely ignoring human common sense or intent.
AI Failures: Fix Bad Data Before Building the Fountain
๐ฑ๐ฎ% ๐บ๐ฎ๐๐ฐ๐ต ๐ฟ๐ฎ๐๐ฒ. Thatโs what an AI portal gave me today for a senior Data & AI leadership role. The justification? "0 years of applicable experience." Context check: Iโve spent the last 15 years leading data and AI initiatives. My immediate thought: ๐๐ช๐ฅ ๐ ๐ข๐ค๐ค๐ช๐ฅ๐ฆ๐ฏ๐ต๐ข๐ญ๐ญ๐บ ๐ถ๐ฑ๐ญ๐ฐ๐ข๐ฅ ๐ข ๐ฑ๐ช๐ค๐ต๐ถ๐ณ๐ฆ ๐ฐ๐ง ๐ฎ๐บ ๐ค๐ข๐ต ๐ช๐ฏ๐ด๐ต๐ฆ๐ข๐ฅ ๐ฐ๐ง ๐ฎ๐บ ๐ณ๐ฆ๐ด๐ถ๐ฎ๐ฆ? I checked the portal. The CV parsed correctly. So I asked their support chatbot. The diagnosis? A classic timing bug. The scoring algorithm triggered before the resume was fully parsed and processed.
๐ Read all writings on LinkedIn โ
A long time ago, in university, I conducted research on cryptography. I still bring that same curious mindset to my daily work.
- ForwardDiffSig: Co-author of The ForwardDiffSig scheme for multicast authentication, published in the IEEE/ACM Transactions on Networking, focused on secure and efficient multicast network authentication.
- Merkle Trees: Co-author of On the Performance and Use of a Space-Efficient Merkle Tree Traversal Algorithm in Real-Time Applications for Wireless and Sensor Networks, published in the IEEE International Conference on Wireless and Mobile Networking and Communications, focused on optimal techniques to traverse Merkle Trees.
- Forward Secure Signatures: Co-author of An Optimized Double Cache Technique for Efficient Use of Forward-secure Signature Schemes, published in the PDP2008: 16th Euromicro Conference on Parallel, Distributed and Network-Based Processing, focused on making Signature Schemes with Forward Security usable. Forward Security is the property by which the signature remains valid even if the private key gets compromised.