I’m an engineer who builds AI systems, and I take on consulting work. I’ll help you figure out where AI actually fits — then build the pipelines, chatbots, and agents that make it real. No hype, no jargon, no sales dance.
Most AI work right now is sold on hype and delivered on hope. I’d rather be specific, honest, and willing to say “don’t build this”. It’s me you’d be working with — one engineer, no account manager in between — and I’ll write your roadmap in plain English before you owe me anything.
Most teams have heard the AI pitch a hundred times. What they need is an honest look at their business, an inventory of the tasks AI is actually good at today, and a roadmap that starts small. We sit with you, listen, and write the plan down in plain English.
A pair of short conversations where we learn what actually happens in your business day-to-day. No pitch deck, no jargon test.
We write up where AI genuinely helps, where it doesn't, and where we'd start. You get a roadmap in plain English that's useful even if you never hire us again.
We pick one problem, ship a working version in weeks, and put it in front of the people who'll actually use it. We iterate with real feedback, not committee.
Software isn't done when it ships. We stay on, watch the dashboards, retrain when needed, and make sure the thing keeps earning its keep.
Systems where a model is the engine rather than a chat box. These are my own projects rather than client engagements — I’m writing them up properly elsewhere, and this list will grow as I do. Enough here to tell whether we’d work well together; happy to go line-by-line on a call.
Between rounds the model reads how you're actually playing, plans the next one in natural language, and emits it in a grammar the engine executes. Every output clears an allowlist before it touches the simulation, and the route caps tokens, time, body size and memory — it can't be talked into a bigger bill. Pull the API key and it falls back to a deterministic director and still plays.
A planner delegates to a bounded pool of workers, and each one iterates until an independent verifier passes — not until the model says it's finished. That distinction is the whole thing. Underneath: test-execution verifiers, a convergence tracker for stalled loops, and a budget guard that ends a run early rather than overrunning spend.
A purpose-built corpus queried through a CLI, so every answer traces back to something you can re-run rather than something the model recalled. Data-gathering is separated from interpretation, and an adversary pass tries to destroy each claim before it's recorded. Its first run killed three of five claims that ordinary self-review had waved through — which is the argument for building it this way.
It’s running on the multi-tenant platform itself, not a mockup: grounded on a per-tenant knowledge base, rate-limited, hardened against prompt injection, and built to say “I don’t know” instead of guessing. Try to catch it overstating something — that’s the useful test.
Hi — I'm the Sol Computing assistant. Ask me about what we do, our process, or how an engagement usually goes. I'll keep it plain-English.
Don't see yours? Send us the question — we'll answer by email before we'd try to sell you anything.
Two free conversations. A plain-English plan. No pressure, no sales dance.