I build the systems that trace, screen and trade on-chain.

Token forensics and contract analysis bot

Send it a contract address, get back a full structured breakdown in seconds: sniper and insider wallets, team allocation, top 10 and top 70 holder concentration, bubblemap cluster detection, the share of holders funded straight from a centralised exchange, holder count and average bag, and a scam or legitimate verdict. Output is formatted to be posted publicly, so the analysis doubles as content.

What it proves

I systematised my own investigative method into something repeatable. The judgement came first, the automation second, which is the right order and the rarer one.

How others use it

Retail buyers screening before they enter. Analysts and KOLs producing breakdowns at speed. Desks and funds using it as first-pass triage before a human looks.

Python · Telegram Bot API · On-chain holder and transfer data · Cluster detection · Automated report formatting

Meteora DLMM automated exit bot

An autonomous exit system for one-sided SOL liquidity positions on Solana. It auto-discovers every open position across pools, caches candles per pool with staggered fetches to stay inside free-tier rate limits, screens mint and freeze authority on-chain for rug risk, arms itself only after entry conditions settle, and closes on a strict signal confluence evaluated on closed candles. Priority fees are set dynamically from live network sampling. Every exit writes itself to a trade log with entry value, exit value, PnL, hold time, unclaimed fees and transaction signatures. A separate pre-entry scorer grades any pool before capital touches it.

What it proves

Engineering under real financial consequence, and the willingness to be wrong in public. I killed an exit signal mid-build after realising it fired at the start of every dump, which was structurally backwards for a strategy where the dump is the fee-earning event. Reasoning about mechanism beats pattern-matching indicators.

How others use it

Liquidity providers farming volatile pairs who need mechanical exits rather than nerve. The pre-entry scorer stands alone as a pool quality filter for anyone sizing a position.

Node.js · Solana · Meteora DLMM SDK · GeckoTerminal · DexScreener · pm2 · Telegram alerts · Supertrend, RSI, Bollinger, MACD

Pump.fun launch data pipeline and strategy backtester

It started with screen recordings of a streamer who claimed big daily profits. I pulled frames from the video, read his terminal with OCR and identified his wallets, and watching them showed his realised PnL falling during both sessions. So I stopped guessing and built the instruments. One Python collector subscribed to every pump.fun launch, sampled them, and recorded what a trader actually sees in the first minute: buyer count, SOL in, top-wallet share, curve progress, sell pressure and launch mode, then followed each price path out to thirty minutes over RPC. A second collector paper-traded the streamer’s wallets in real time, reading each swap from balance changes so it caught his trades whatever app or venue he used. Both ran unattended for five days with spend tracking, low-balance alerts and recovery after reboots: 73,285 launches and 9,021 complete price paths. I then tested 3,797 entry filters against 80 exit strategies each, choosing rules on the early days and checking them on the later ones.

What it proves

That I will pay to find out rather than lose money learning live. None of the 3,797 rules made money, even with zero fees. One rule looked positive until I noticed it used information you could not have had at the moment of buying; rerun honestly, it lost about 6%. Only 7.5% of the active coins ever doubled. I shut the collection down and moved the effort to the side of the trade that earns from dumps, the DLMM bot above.

How others use it

Traders who want a strategy tested against real launch data before they size into it. Funds and launchpads studying first-minute behaviour. The collector schema and the train-then-test discipline carry over to any strategy question where the data is noisy and the incentive to fool yourself is high.

Python · PumpPortal WebSocket · Helius RPC · Jupiter pricing · pandas · Telegram alerts · Out-of-sample backtesting

Trenches X-Ray: a public pump.fun forensics dashboard

The backtester answered one question in private. This one puts the evidence in public, running on Solana data on Dune. It has six sections. A launch funnel shows how many tokens launch each day, how few graduate and how much of it comes from serial devs. Block-0 bundles finds wallets buying in the same slot the token was created and checks how those launches graduate against clean ones. Dev farms tracks creator wallets launching hundreds of tokens and pulling SOL out of them. Smart early buyers finds wallets that get in during the first ten minutes and actually realise a profit, then shows what they are buying now. A rug radar scores recent launches on simple red flags. Token X-Ray takes any pump.fun mint and shows who holds it, how they got in and who funded them, including clustered supply, fresh-wallet supply and how much the dev and block-0 buyers still hold.

What it proves

The backtester was not a one-week fluke. Over 13 full days in late October, roughly 47,000 tokens launched a day, only 0.33% graduated, and 73% of launches came from serial devs. Anyone can check those numbers, because the queries are public. That is the difference between a claim and evidence.

How others use it

Traders checking a mint before they buy. Researchers and journalists who need real numbers on how the launchpad behaves. Launchpads and exchanges looking at bundling and dev-farm activity on the tokens they list.

Dune SQL · Solana decoded tables · pump.fun program data · Parameterised queries · Funding-cluster tracing

Open the dashboard on Dune →

Crypto Battlefield

A live 3D trading visualisation where market data becomes a fight rather than a candle chart. Buys and sells become units on the field, transaction volume sets the tempo, and the front line moves with price. Built as a single self-contained page on Netlify, rendering with Three.js against a live DexScreener feed, with real-time chat on Firebase, four visual themes, a Token vs Token versus mode, and an embeddable widget so any community can drop it into their own site. Seven major versions shipped so far, with distribution running through X, TikTok and YouTube Shorts.

What it proves

I take products from idea to a live URL with real users, then keep going. Seven versions is the signal, not the first one. It also covers the full stack of the job: build, host, instrument, distribute.

How others use it

Traders who want to feel market action rather than read it. Token communities embedding the widget for engagement. Platforms looking for a retention surface that is not another chart.

Three.js · WebGL · Firebase · DexScreener API · Netlify · GoatCounter · Embeddable iframe widget

X monitoring and reply agent

A Python system that watches around fifteen crypto influencer accounts on a push stream, scores each new post for whether it is worth engaging, analyses it with an LLM, and delivers one recommended reply to Telegram as a tap-to-copy block with a direct link to the exact post. Each watched account carries its own tone profile, analyst or sarcastic, switchable live from Telegram with no restart. Posting to X stays permanently manual by design. 374 tests passing. It exists to replace the cost of a social media manager.

What it proves

Production judgement, not just working code. I cut A/B reply candidates down to one because operator speed mattered more than the feature. I kept a human on the trigger to protect the account. And I cut polling from 5 seconds to 60 after a $10 lesson in API burn, then set a $0.50 a day ceiling. Agents die on unit economics, not on model quality.

How others use it

Any founder or creator replacing a social media hire. More broadly, the pattern generalises: watch a stream, have a model draft, let a human approve in one tap on their phone. That shape covers most agent work that survives contact with reality.

Python · TwitterAPI.io push stream · OpenAI · Telegram Bot API · Per-account prompt profiles · 374 tests

AI estimating workflow

As a former senior cost manager with eight years across consultancy, main contracting and specialist subcontracting, the largest package I priced solo was around £30M. So while estimating at RapidQS, I automated my own job with an AI-assisted workflow: a structured take-off and bill template, a Claude-based rating and checking layer, Kreo integration for measurement, and a QA process that measurably cut client revisions. It is the reason I can turn around a bill of quantities in a fraction of the usual time without the accuracy loss that normally comes with speed.

What it proves

The thing most AI projects lack: a domain expert who knows exactly where the model is allowed to be trusted and where it must be checked. The QA gate is the product, not the generation step. This is the same discipline that makes an LLM deployment survive in any regulated or high-consequence industry.

How others use it

Contractors and consultancies pricing faster without adding headcount. More generally it is a template for applying AI inside any expert profession: encode the method, automate the mechanical part, gate the output, keep the expert accountable.

Claude API · Structured prompt workflow · Kreo · Excel estimating templates · NRM2 · Multi-stage QA review

AI coaching intake and plan engine

A nineteen-question conditional intake form feeding an automation layer into the Claude API, which returns a complete personalised training programme, nutrition plan and diabetes management protocol, formatted and delivered to the client. Three subscription tiers sit behind it. The prompt is the product: it carries the coaching methodology, the safety boundaries, and the structure of the output, so the quality holds without me in the room.

What it proves

End-to-end AI product delivery with no engineering team, and the harder half of that job. Turning a professional's judgement into a prompt that holds up on inputs the author never anticipated is most of what commercial LLM deployment actually is.

How others use it

Coaches, clinics and any expert practice that needs to serve more people than hours allow. The architecture is domain-agnostic: conditional intake, automation middleware, LLM generation, branded document out.

Claude API · Typeform Pro · Make · Prompt engineering · MailerLite · Stripe · Tiered subscription delivery

Coaching guides and publication pipeline

A library of branded guides and long-form drafts, produced through a repeatable pipeline rather than by hand: programmatic document generation, conversion to print-ready PDF, structural validation, and a visual check where every page is rendered to an image and inspected before anything reaches a client. English is always the locked master, reviewed before any Arabic adaptation begins, so the two versions never drift apart.

What it proves

The unglamorous discipline almost nobody applies to generated output. I built verification into the pipeline instead of trusting the generator. Bilingual delivery with a locked source of truth is the same problem as versioned content at scale.

How others use it

Creators and educators shipping premium digital products in more than one language. The toolchain itself is reusable for any documentation, report or template output that has to look correct, not just parse correctly.

Node.js · docx · WeasyPrint · LibreOffice conversion · Structural validation · Page-to-image QA · Arabic RTL typesetting