SynthworxStart a strategy session
Menu
WorkServices & feesAutomationProductsNotesAboutContactStart a strategy session — $1,500

Selected work

Four systems. Measured.

One built at Amazon, two independent products, one research project. Each case study follows the same structure: context, problem, architecture, what I did, results, and what it proves.

Amazon work is sanitized: no confidential code, customer data, or internal systems. The numbers are the published outcomes of the work, not a forecast for yours.

Built at Amazon2015–2022

Running LLM systems at 2M requests per hour

A sanitized account of the systems, tradeoffs, and operating discipline behind a production ML platform serving more than two million requests an hour at 99.99% availability.

99.99%availability at 2M+ requests/hour
Read the case study→
Independent R&D2026

A 1.1M-token memory on an 8 GB laptop

Kamaloki keeps a compact, probeable KV store of a long agent session and splices the relevant spans back into a small model's working context. Planted-fact recall matched a frontier model reading the raw session — at 4.93 GB.

15/15planted facts recalled from 1.1M tokens
Read the case study→
Independent product2025–2026

A browser-native agent with 3–10× token savings

Tappi Browser is an Electron-based AI browser with a browser-native agent: referenced element indexing, domain playbooks, and persistent learning. The result is 3–10× fewer tokens than conventional browser tooling, with no navigator.webdriver footprint.

3–10×fewer tokens than browser-tool baselines
Read the case study→
Independent product2026

Pikki: a live, paid Chrome extension

Pikki is a live product with three paid plans, a local-first data model, and a self-updating native helper. It checks postings for ghost-job and scam signals, finds referral paths, and prepares applications from a verified profile.

Livepaid product · $19 / $39 / $79 plans
Read the case study→

More proof

Other systems worth naming.

Products
01Tappi for ChromeIndependent product workMulti-harness AI sidebar with native messaging and Office integration. ~36.5K LOC, 69 test files.~36.5K LOC02AI Tutor (Android / Fire OS)Client / independent product workRealtime voice tutoring, encrypted data, signed PDFs, GMS-free. 632 Kotlin files, ~95K LOC, 522 tests at checkpoint.522 tests03Tappi Python / JSOpen-source utilityMIT browser-control utilities for AI agents. Public, documented, and used in the Tappi stack.MIT

Your system

Do you recognise the failure?

If your AI feature is slow, expensive, unreliable, or right just often enough to be dangerous, start with the 90-minute session. It produces a one-page memo and three prioritized recommendations, and the fee is credited toward the two-week diagnostic.

Book the session — $1,500