// PROOF
The cut decides, not the carat.
Two stones of identical mass differ in value by an order of magnitude, because light only returns if the geometry is right. Same models, same data, same headcount — the architecture decides whether anything comes back out.
Every row below is a named system with a number and a note on how the number was measured. Each has its own URL, so any single one can be sent on its own.
- Decima · Open Source · 2026
A 122M-parameter decision model that never changes its answer when you reorder the options
Decima is a small, calibrated, multilingual decision model — situation, question and free-text options in; one probability per option out — that runs in about 20 ms on a single laptop CPU core and is released with open weights under Apache-2.0.
- ~20 ms per decision on one laptop CPU core (int8, 4 options)
- 122M parameters — 2.6–6× smaller than the open decision models we compared
- 0% answers changed when the options are shuffled (competitors 10–27%)
- 0.764 accuracy on Laya's multilingual protocol (Laya-multilingual 0.607)
- Cortex-Axon-Synapse · Mistix AI · 2024–2025
A three-layer orchestration backbone that cut client deployment time by 75%
A proprietary three-layer AI orchestration backbone — autonomous planning, deterministic NL-to-API execution, and a secure high-throughput gateway — that made every subsequent client deployment a configuration exercise rather than a rebuild.
- 75% reduction in client deployment time
- $1.5M+ B2B enterprise contracts secured and delivered
- Oasis · Smartway Solutions · 2025–present
Natural-language query over multi-gigabyte files, for people who do not write SQL
A cost-cutting engine that let Sales, Marketing and Operations interrogate multi-gigabyte data files in plain language, bypassing the business-intelligence queue entirely.
- 10x compression of end-to-end processing and documentation time
- 0 BI tickets required to answer a routine data question
- Confidential · 2023–2024
A six-step multi-agent pipeline that classified soil 85% faster than the manual process
Fully automated USCS geotechnical soil classification from raw borehole data using a six-step multi-agent pipeline with MCP-based tool integration, replacing a slow and inconsistent manual reading process.
- 85% reduction in processing time per borehole log
- Confidential · 2023–2024
Cutting monthly cloud GPU spend by 60% with PEFT and quantization
Applied parameter-efficient fine-tuning and aggressive model quantization to move production inference onto constrained hardware, removing the majority of a recurring cloud GPU bill without a quality regression.
- 60% reduction in monthly cloud GPU expenditure
- Confidential · 2023–2024
An LLM-as-a-Judge eval framework that tripled deployment frequency
Instituted continuous automated evaluation of model outputs, replacing ad-hoc manual review as the release gate and increasing how often the team could safely ship.
- 300% increase in model deployment frequency
- Confidential · 2022–2023
A hybrid-retrieval RAG assistant that removed half the human support load
A multi-layered retrieval-augmented assistant with hybrid semantic and lexical routing, which halved inbound support tickets requiring a human and materially improved retrieval precision.
- 50% reduction in human support tickets
- 35% improvement in retrieval precision
- Confidential · 2022–2023
An async backend holding 10,000 concurrent daily tasks at 99.9% uptime
A Celery and Redis task backend that reliably absorbed upwards of ten thousand concurrent daily task requests, giving the AI services above it somewhere safe to fail.
- 10,000 concurrent daily task requests handled
- 99.9% uptime
- Freelance · 2019–2022
ELT pipelines at 99% data integrity, and 70% of manual validation removed
Apache NiFi and Airbyte ingestion pipelines with machine-learning validation models, which held data integrity at 99% while automating away the majority of manual validation hours.
- 99% data integrity rate across ingested records
- 70% of manual data validation hours automated away
- Smartway Solutions · 2025–present
Dismantling Jira for a Forward Deployed Engineer model, at 10x output
Replaced traditional ticket-driven project management with a Palantir-style Forward Deployed Engineer model, embedding engineers directly in client problem spaces and framing high-velocity pods under The Harvest Methodology.
- 10x output in a hypersonic-speed delivery environment