I build systems that turn observations into products.
I am a software engineer, AI builder, and startup co-founder. I investigate user friction constraints, formulate validation models, design secure edge architectures, and deploy scalable products. Technology is secondary to discovering actual problems worth solving.
How many new active users are organically referred by each current active user per cycle.
The percentage of existing users who drop off the platform at the end of each cycle.
Initial capital spent to seed the initial user population and kickstart the loop.
Classified Products
On-Device Visual Assist Companion
Build an ultra-fast visual interpreter for blind users that bypasses iOS native install barriers.
VoiceOver delays and cloud-roundtrip latencies (2-4 seconds) cause low-vision users to drop off assistive platforms quickly.
Implemented a browser-native inference hierarchy: Tier 1 executes object classification locally in WebGL via TensorFlow.js COCO-SSD (80-120ms); Tier 2 processes document OCR on-device; Tier 3 routes only complex visual queries to Gemini 1.5 Flash. Bypassed iOS browser SpeechSynthesis autoplay gates using a cyclic user-triggered dummy audio context.
Capability-Gated Agent Proxy
Enforce strict least-privilege capability gates on autonomous coding agents (Claude SDK) inside VS Code workspaces.
Prompting directives are easily subverted or ignored. Coding agents need hardware/process level security boundaries.
Built a security middleware proxy in TypeScript. The system intercepts agent tool executions (e.g. file writes, shell tasks) before process evaluation, mapping calls against a strict schema. Flagged executions prompt the user with interactive approval webviews.
Real-Time Voice Dubbing & Quantum Matcher
Optimize compute budgets of heavy PyTorch DSP separator architectures.
Offloading multi-track Librosa/Demucs computations to un-throttled Cloud Run instances solves CPU serverless time limits.
Deployed FastAPI containers with dedicated PyTorch runtimes. Integrated a simulated 4-qubit quantum variational circuit using PennyLane to evaluate mood-based matching vectors, validating forward backpropagation gradients across hybrid networks.
Bilateral progressive disclosure matcher
Design a privacy-preserving professional network that gates identity leakages.
Conducted 50+ interviews with founders; found that progressive data reveal states (bio ➔ credentials ➔ contact) built 80% higher trust.
Engineered using Next.js 16 and React 19. Set up schema validation blocks with React Hook Form and Zod resolvers to prevent data leaks. Deployed on high-frequency Vercel branch deployment pipelines.
Capabilities & Systems Stack
A look into the technologies, platforms, and strategies I leverage to design pipelines and ship products.
The Future Radar
A circular radar tracking and evaluating technology domains that will frame the operating layer of early-stage SaaS networks.
Future Discoveries
People no longer search. They delegate.
Interfaces are disappearing. Reasoning is replacing navigation.
Every industry will eventually have an AI operating layer.
Observations
Most startups fail before writing code.
Prompt security gating is a leaky model.
Build Log
Validated iOS VoiceOver lag constraints; adjusted edge-inference classifications.
Deleted heavy cloud TTS APIs, migrating 100% of speech rendering on-device.
Enforced 4-qubit simulated Pennylane backpropagation logic across audio recommendation stems.
Mind Map Graph
Hover coordinates to reorganize the network forces of the operating system.