I build software that makes complex systems easier to trust.

MandateFlow places a policy gateway between an agent and protected tools, checking the provenance of the information behind an action.

A support request can become a CRM case when its lineage is allowed. A payment-failure request can look similar while carrying a different data boundary.

Same-looking request. Different lineage. Different outcome.

Constraint

The same public call can arrive with different references, so permission alone cannot explain whether the next tool call is safe.

My role

Team project. This page keeps my individual contribution at the system level until the final scope is documented.

Decision

Use a Go MCP gateway to resolve provenance, apply policy, and write a redacted receipt before the protected tool runs.

Evidence

Support → CRM is allowed; payment → CRM is denied before execution, with an aggregate payment-failure path that does not re-identify a customer.

• Go
• MCP gateway
• React / Fastify
• SQLite receipts

Apple Health imports started as a parsing problem. Profiling showed a different shape: Python object creation and per-record work were holding the file hostage.

The documented test used a 626 MB export with 1.35 million records, workouts, and activity summaries.

Constraint

A six-month export took about 26 minutes, and the bottleneck lived in the ingestion architecture rather than one slow database query.

My role

I profiled the importer, corrected the bottleneck hypothesis, and rewrote the hot path around memory-mapped byte scans and parallel workers.

Decision

Skip the XML tree for record extraction, emit Parquet shards, and let DuckDB assemble the results after the expensive work is parallelized.

Evidence

Reported 26 min → 9.77 s in the documented test environment. The result is a benchmark, not an unqualified production SLA.

• Python
• mmap
• Parquet
• DuckDB
• ProcessPoolExecutor

Screenshot Declutterer turns a desktop full of unnamed captures into a local keep / unsorted / trash workflow, with optional on-device rename suggestions.

The useful part is small: recognize what a screenshot is, make the filename searchable, and let the person decide what happens to the file.

A timestamp becomes a useful filename while the image stays local.

Constraint

Uploading screenshots just to classify or rename them would defeat the privacy premise. Permanent deletion would make a mistaken decision expensive.

My role

I built the local triage flow and evaluated the vision rename path on a 16 GB Apple Silicon Mac.

Decision

Use LiteRT-LM with Gemma 4 E2B locally, and move discarded files to the native Trash through send2trash rather than deleting them permanently.

Evidence

Across 39 screenshots, LiteRT-LM averaged 6.5 s per image versus 20.9 s with Ollama; the model footprint fell from 7.2 GB to 2.4 GB.

• Python
• Flask
• LiteRT-LM
• Gemma 4 E2B

Fitback connects two user photos and retailer-specific sizing data to a recommendation that includes a person’s fit preference.

Fashion sizing is difficult because a generic chart cannot see the difference between a standard, tight, or loose fit preference.

Constraint

The system had to turn messy sizing charts and body-pose input into something practical enough for a Shopify purchase flow.

My role

Technical project lead / self-employed. The final split across product, ML, frontend, backend, and team leadership is kept broad here until confirmed.

Decision

Combine pose validation and 17 keypoints with an MNASNet + MLP recommendation path, and use AWS Textract to extract retailer chart data.

Evidence

LinkedIn reports approximately 2 cm mean absolute error and 0.1 s inference. Fitback won the DesignAI category at InspireCon 2025 and was covered by The Business Times.

• Shopify Remix
• AWS Textract
• MoveNet / PoseNet
• MNASNet + MLP

At Primech, the work moved between robot dashboards, REST APIs, authentication, IoT sensor data, internal applications, and documentation.

The systems were already in motion. The job was to improve the next change without losing the context held by the existing codebase.

Constraint

Legacy Flutter and PHP tools, real-time widgets, and authentication issues made small changes depend on understanding more than one layer.

My role

I worked across dashboard CRUD APIs and UI, JWT authentication issues, JavaScript refactoring, Python IoT services, Flutter, PHP, and architecture documentation.

Decision

Treat documentation and system boundaries as part of the implementation, so the next person can trace a request from the interface to the device.

Evidence

The two R&D internships covered robot dashboard work, a Python IoT backend, legacy-code modernization, internal tooling, and older-system documentation.

• JavaScript
• REST APIs
• JWT
• Python IoT
• Flutter / PHP