Mechatronics engineer turned hands-off systems builder. I cut an industrial process by a conservative 85%. Now I build falsification engines, agent workflows, operating tools, and the cited AI twin on this page.
The short version, before the systems: what I do, where it came from, and the one result that was measured inside somebody else’s company.
Since mid-2025 I have been running Snedker Automations as a Forward Deployed Engineer, while building Popper and other autonomous trading-research systems.
I have lived in five countries over 34 years: 26 in Australia, then New Zealand, two years in Canada, most of a year in Mexico, and five in Germany. From each base I travelled outward — 30+ countries visited, the US from Canada, Guatemala from Mexico, western Europe from Stuttgart. Four off-grid builds, and a Canadian business operated from Mexico. That range is not decoration. It is why I am comfortable entering a system I do not know, watching before touching it, and finding the constraint others have normalised.
What that looks like in practice. At ThyssenKrupp I let three projects go past before I built anything. I asked why cable management took two engineers five to six weeks, found the constraint, and built the replacement with the older engineer who would have to live with it. Project four ran with one engineer in about two weeks. Nobody had to be talked into using it.
That was 2017. It is still the cleanest example I have, because it is the one that was measured inside someone else's operation — and that is the specific thing a personal portfolio cannot fake. What changed since is the domain, not the sequence: Popper exists because five years of trading told me which claim needed falsifying, and the twin on this page exists because a CV cannot be interrogated. Same behaviour, harder problems. The current work is in the atlas below.
Nine systems, ordered by what they prove rather than by how new they are. Problem, architecture, evidence, limitation — you should not need to open a card to understand the work, and opening one should tell you considerably more than a bullet list. Behind these sit roughly seventy workflows built and about seven that reached production, all of them for Jake, his family, or a friend’s business.
Most research tooling helps produce an attractive backtest. Popper assumes the strategy is wrong and tries to kill it before money is involved. The rule it enforces is the one that cost Jake a live account: a result that only survives one split, one period, or one parameter value is not a result.
A Director loop coordinates three stateless LLM functions over a shared Alpha × Wrapper × Context language. Generation is the only place a model is trusted; every accept-or-reject decision is a deterministic scorecard. Survivors carry their own falsification record into portfolio construction, and rejects become memory so the same dead idea is not re-proposed.
Runs as a fleet: three operators across four machines on a shared heartbeat table, so a run continues when any one box sleeps. Cost is held near zero by stacking subscription-quota runs rather than paying per token.
The core loop runs end to end. The last solid bulk measurement was about one generated alpha every 2.7 minutes. The adjacent public framework shows the validation methods — walk-forward, Monte Carlo, deflated Sharpe, probability of backtest overfitting — without exposing the private platform.
No Popper-discovered strategy is claimed profitable or live. The honest state is a working research engine, not a proven edge. Saying otherwise would repeat exactly the mistake it was built to prevent.
Several artists share one stall and one card reader. At the end of a market someone has to work out who sold what, and who owes what share of the card fees. Done on paper it is slow, done in a spreadsheet at the end of a cold day it is wrong, and the person holding the reader silently absorbs everyone else's transaction costs.
One Google Apps Script web app deployment serves every market, selected by a ?m=<sheetId> parameter rather than a copy per market — so a fix ships everywhere at once instead of being re-pasted into forks. A Data Entry › New market dialog spins up a whole market from a name and a list of artists: sheet, tabs, and vendor columns.
SumUp's 1.39% card fee is split per line into Fee and Net columns at entry time, not reconciled afterwards. Each vendor is paid their own net, so the fee lands on the sale that incurred it. That single column is the difference between the app being used and being abandoned.
Live and in use across multiple markets, with two to three vendors on a stall on the busier days, including a real trading weekend in August 2026. A portfolio-safe version of the allocation logic is public.
Real users, small scale. This is a family business and the stallholders beside it — not a customer with a contract, a support rota, or a load profile. It proves adoption and correctness under real money, not throughput.
A craft business needs a consistent posting cadence and has no time to produce one. Generic AI captions solve the time problem and destroy the thing people actually follow the account for, which is its voice.
Image from Google Drive → caption generated against the account's own tone → row lands in an Airtable approval queue → a daily check picks up approved rows → spacing logic decides whether today is a sensible day to post → publish through the Instagram Graph API. Google OAuth2 and long-lived Graph tokens both refresh on their own.
The approval queue is the whole point. The system is allowed to do the labour and not the judgement: nothing reaches the account that a human has not looked at. That is the same gate Job Hunter uses on email, and the same reason Popper never lets a model score its own idea.
Jake's first workflow to reach production and still the longest-running one. Years of uptime through token expiries and platform changes is an adoption number, not a demo — it survived neglect, which is the harder test.
One account, one business, no external SLA. It shows a trust boundary designed correctly and left alone for years; it does not show multi-tenant scale.
Short-form content for a venue product needs volume and cannot look generated. Every cheap pipeline produces the same drifting stock footage with captions floating half a second off the audio, and viewers have learned to scroll past it in under a second.
Remotion renders video as React, so a reel is a component and a config file rather than a timeline someone drags by hand. Narration comes from edge-tts with word-level timings, and those timings drive the caption track directly — the words land on the syllable because they are the same data, not because someone nudged them.
The mechanism works: render from config, narration with word-level timings, captions driven by those same timings. The output does not. It still reads as generated filler, which is the whole problem this was meant to solve, so I am not going to claim otherwise on a page about hiring me. An earlier attempt through a generic content tool was abandoned for the same reason.
This is the one system on the page that has not met its own bar. It is here because the honest portfolio includes the thing that is not working yet, and because the gap is instructive: the engineering was never the hard part, taste and source material are. Next build.
Running eight builds at once, the expensive failure is not slow work — it is a project quietly stalling on a blocker nobody has looked at for three weeks, while attention goes to whichever one was touched most recently.
Next.js front end over a FastAPI service and a single state file per project. Each project answers exactly three fields: next action, progress, bottleneck. The constraint is deliberate — anything richer becomes a second project to maintain.
A hybrid brain: Claude reads the state and proposes the next move, Jake accepts or rewrites it, and only the approved version is written back. The model never silently edits the thing being used to make decisions.
Single-user, local, started on demand. It is the control surface for one operator's estate, not a team tool, and it has never been tested with a second person writing to it.
A strategy validated in one engine can mean something subtly different once it is hand-written again for TradingView or MetaTrader — and the disagreement usually surfaces as live behaviour that does not match the backtest, which is the most expensive place to find it.
Deliberately subordinate to Popper. Popper owns the grammar, the validation and the feature semantics as the single source of truth; Forge owns nothing but the two renderers. Adding a third target means writing a renderer, never a second dialect.
Coverage guards enumerate the shared feature set and fail the build when either target lacks a renderer for something the DSL can express. Unsupported exit logic is refused outright rather than approximated — an approximation is exactly the silent drift the tool exists to prevent.
Written end to end by an autonomous agent run, then reviewed. That is a demonstration of Jake's current working method rather than a claim about it.
The plain-English front end depends on a local Claude CLI, so it is a deployment surface for Jake rather than a hosted product. Built, and honestly low-use.
Foundation, entry refinement, time and context, trade management, then exits and risk — each compared against the correct reference, baseline early and the winning system later.
Gates between stages stop weak paths before more parameters are added, because the cheapest way to fake an edge is to keep adding degrees of freedom to a foundation that never had one.
The associated bot passed both phases of a real prop-firm challenge and reached a live account.
It then lost that live account before a payout. No automated-trading profit has been taken. Those two facts always travel together — and the second one is why Popper exists.
Good roles are crowded before a conventional application reaches the person who owns the problem.
A scheduled run drives Claude through explicit search lanes, finds the decision-maker behind a signal, and writes a short Gmail draft. A local dashboard and a JSON ledger handle dedupe and outcomes. Sending stays manual by design.
Gmail audit: 15 sent, 12 delivered, 3 invalid-address bounces, one automatic acknowledgement, zero human replies. Nine records were mislabelled as draft. Reported here as measured, not as hoped.
The verification gate and ledger are both broken, so it is paused in draft-only mode until they and the AI-written subject lines are fixed. It is on this page as an honest in-progress system.
Exact-origin browser gateway → Supabase Edge for rate limiting, sessions and idempotency → retrieval over a public-only corpus → grounded generation on the VPS → an allow-list that discards any citation the model invented.
When vector search returns nothing usable it falls through to Postgres full-text rather than answering ungrounded, and the whole path fits inside 23.5 seconds of a 25-second budget. Groq fails over to NVIDIA rather than to an apology.
It can quote the record and cannot yet hand you an artifact. Ask it something it has no evidence for and it will say so rather than improvise — which is the intended behaviour and also the shortest demo of the point.
No fake dashboard shots. These are sanitized outputs from systems Jake built. The quant chart uses synthetic demo data and is not evidence of trading performance.
Published work / verifiable without asking me
Three books, 150+ episodes and two courses to about 65 students — all of it about trading and trading psychology, written while doing it full time. It is a technical-education track record, not an active revenue stream, and I would not present it as one.
Read it backwards, from now to the start. The domains change every few years and the behaviour does not: observe first, find the real constraint, build it with the person who has to use it, then keep it running.
Fifteen months of parental leave spent learning n8n properly, then AI agents, then moving to CLI-first building. I launched the business in September 2025. The main chatbot reached roughly 90% and never shipped, because latency and booking integrity were not good enough to put in front of someone’s customers.
Five years of full-time discretionary trading, returning 5,800 USD in lifetime payouts. Not a success story dressed as one — it was an apprenticeship in risk, execution, MQL5 and PineScript, live teaching three times a week, and exactly why an attractive backtest proves nothing.
In May 2026 my wife and I made the call: discretionary trading dropped to a hobby, and the algorithmic work went the other way, from side project to near-daily. That decision is what started this job search — not a lack of options.
Remote team lead for 12 core contributors plus around eight part-timers: I planned the experiments, distributed the work, trained the group, analysed the results, and made the call at the end. The company closed in November 2023 for reasons unrelated to my project. Teaching returned about 19,000 AUD across two separate ventures: roughly 9,000 from two courses to about 65 students here, and just under 10,000 from Zen Traders Academy later.
I sold solar door-to-door for EV Green in Kelowna — lead gathering, booked appointments, daily presentations, closing, and handling payment — while running a rickshaw business and building a school bus out in the evenings.
The bus became an Airbnb at 72% occupancy from March to November, which I operated remotely from Mexico until a land-control problem forced me to cancel every booking. That is the whole “run it from anywhere” thesis, and also how it ends when you do not control the ground underneath it.
Forklift and tractor work in kiwifruit orchards paid for the conversion I was building alongside it. I bought the van for about 2,000 NZD, put in roughly 3,000, and sold it for about 11,000.
Three projects went by before I built anything. I watched, asked why cable management took two engineers five to six weeks plus frequent help from a third, and found the bottleneck.
I built the database with the older engineer who would end up using it. The fourth project ran with one engineer and the tool, in about two weeks. Nobody resisted it — they were relieved to get back to the engineering they actually liked. I was offered a two-year contract and turned it down to go travelling.
The measurements are mine, not audited company figures. I use 85% because it is the conservative end.
B.Sc. Mechatronic Engineering with Honours, and a thesis on a navigation system for pedicle screw alignment in spinal surgery. Six years of hospital work across Theatre, ICU, Dialysis and Radiology, where the cost of getting something wrong is not a rolled-back deploy.
Through all seven of those years I co-owned a window-cleaning business: invoicing, stock, bookkeeping and chasing receivables. The first thing I ever ran.
I have trekked with silverback gorillas, jumped out of planes, dived through Asia, built homes from bare shells, and committed to destinations before I knew the route. None of that made me comfortable with ambiguity. I was already that person, and it is the same trait that lets me walk into an operation I do not know and watch before touching anything.
Here so you can stop reading if one of them is genuinely disqualifying for your role. None of them is a surprise to me, and each one sits next to something I have proved instead.
I have not operated a system under substantial third-party load, against an SLA, or through an on-call rotation. My production users have been me, family, or a friend's business.
What I do have: systems those people still depend on years later, and the judgement to know which of my own work was not safe to ship.
I architect and read Python systems well, but I would do poorly in a timed, no-AI algorithm test. I hand-wrote MQL5 and PineScript for roughly three years; Python has always been AI-assisted.
What I do have: three years of hand-written strategy code before AI existed to lean on, and a working method built around the tools the job actually uses.
My trading work is entrepreneurial and self-funded. It includes real prop-firm evaluations and real losses, not an institutional desk or a profitable automated track record.
What I do have: five years of my own money on the line, which is where the instinct to falsify a good-looking result comes from.
Git discipline and formal testing became serious recently. I bring unusual system ownership and breadth, not five years inside one mature software organisation.
What I do have: end-to-end ownership of everything on this page — problem, architecture, build, adoption, and keeping it running.
The live twin uses a deliberately small recruiter corpus, shows what it cited, and says when the evidence is not enough. In reviewed-demo mode, every sample answer is visibly labelled and links to this page.
“What is the strongest proof Jake can deliver?”
“Tell me the honest trading record.”
“Where is he under-experienced?”
Open the twin ↗I am looking for a Forward Deployed Engineer, Quant Developer, AI Engineer, or AI Researcher role with a real problem and enough ownership to solve it properly.
jakesnedker@gmail.com ↗Same record, three ways in. Take whichever argues your role.
Prefer to interrogate it instead? Ask the twin ↗