Telegram bots, websites and web apps, AI that answers from your own documents, and video and data pipelines, for small teams and organisations in Singapore and Indonesia. Built inside the apps your team already uses, engineered to Singapore company standards, and looked after once they're live.
Click any project for the full story. Most of these belong to private organisations, so the screens here are recreations with sample data. The descriptions and numbers are real.
Telegram bots and websites that run a community organisation's admin.
A community organisation ran its money, attendance, results and outreach through group chats and spreadsheets. I replaced the manual parts with Telegram bots that read what people post, and websites that give leaders the full picture. People keep messaging the way they always have. Within minutes, each report shows up on the right website as a card, a record or a ledger line, and the reports write themselves.
Reads receipt photos and PDFs with GPT-4o, matches them to claims and writes the treasurer's reports. Monthly paperwork went from 2 days to about 10 minutes.
Treasurers mark claims as paid on the web, and the bot's ledger updates itself.
Turns casual chat reports into clean records and live summaries. Before, there was no attendance history at all. Now every week is kept and can be compared over time.
Member cards, groups on a drag-and-drop board, and a live attendance calendar fed by the bots and the leaders' chat. Exports to the same Excel layout the team used before.
Records score reports posted in chat and works out pass rates for each class.
Class calendar, rosters and progress reports, kept in step with each class's Telegram topic.
Reads reports posted in Telegram, Word files included, then creates or updates the right card and replies with its link.
Every lead as a card with days in each stage, nightly stats, and a map of which places work best.
Nightly AI summaries of what's working, with names and personal details removed before anything reaches the AI, plus spend tracking for every app.
Python · Telegram · FastAPI · HTMX · React · PostgreSQL · Supabase · SQLite · OpenAI · Docker · Caddy · nginx · DigitalOcean
Want something like this for your team?
Chat on WhatsAppZoom recordings edited into English and Indonesian, published after one tap.
Long sessions recorded on Zoom with a live interpreter used to be edited by hand, in two languages. Now the pipeline downloads the recording and the interpreter's track, cleans the audio with AI, and cuts an English and an Indonesian edit. A person taps approve, and it publishes both to Vimeo and a backup server, then checks that every file arrived intact.
Finds the new recording and pulls the main and interpreter audio tracks automatically.
Removes noise and echo from the interpreter channel so both versions sound clear.
Cuts and renders the English and Indonesian versions in 1080p, ready to publish.
Uploads to Vimeo and a backup server, then checks every file before calling the job done.
AI cuts the speaker out of the frame so the slides behind them can be swapped for clean ones.
Separates overlapping speakers and lines up dubbed audio with the video.
Python · Zoom API · Vimeo API · ffmpeg · PyTorch · Whisper · OpenCV
Want something like this for your team?
Chat on WhatsAppAnswers from your own documents, with the page each answer came from.
Ask a question in plain words and get an answer from your own documents, with the exact page it came from. When the evidence isn't there, it says so instead of making something up. TradeIQ, my open-source version for company filings, is public on GitHub.
Two new risk factors appear this year: dependence on a single overseas supplier for its main component [1], and new data-privacy rules in two markets that could raise compliance costs [2]. The rest of the section is unchanged from last year [3].
Answers questions from SEC filings and earnings reports, with page references for every claim.
Finds notes by meaning across a large personal library, inside a Telegram bot and phone app.
Reads a job listing and rewrites a CV and cover letter for it, approved with one tap in Telegram.
Python · hybrid keyword and vector search · reranking · pgvector · OpenAI · Claude API · Docker
Want something like this for your team?
Chat on WhatsAppBacktests, sizes and re-weights trading strategies every night.
A system that backtests strategies, sizes every position inside hard risk limits and re-weights itself each night with a machine-learning model. It runs on its own server through a real broker's API, paper trading for now, with a desktop control panel to watch it.
Replays 16 years of market data against every strategy before it touches money.
Position sizing and exposure caps that every trade has to pass.
A machine-learning model re-weights the strategies each night from recent results.
A desktop app showing the bot, server health and costs in one window.
Python · FastAPI · PostgreSQL · scikit-learn · broker API · Docker · Electron
Want something like this for your team?
Chat on WhatsAppValuation and classification models from company and research work.
Before freelancing, I built machine-learning models for a living. At Real Estate Analytics (8PROP) I trained models that value Singapore HDB flats and condominiums to a 3% median error, and deployed models on AWS Lambda that check and clean data as it flows in. At A*STAR I built classifiers that sort candidate drug molecules with an 85% average F1 score. The same skills now run my trading model and video AI.
XGBoost, support vector and neural-network models that value Singapore HDB flats and condos to a 3% median error.
Models deployed inside the data pipeline that catch and clean bad records before they reach reports.
Research at A*STAR: Random Forest, XGBoost, SVM and PyTorch models tuned with Optuna, at an 85% average F1 score.
Re-weights trading strategies every night from recent results, inside hard risk limits.
Separates overlapping voices, recognises who is speaking and cuts a person out of the frame.
A DeBERTa classifier in TradeIQ that blocks attacks on the AI before a question reaches it.
Python · scikit-learn · XGBoost · PyTorch · TensorFlow · Optuna · AWS Lambda · SQL · Airflow
Want something like this for your team?
Chat on WhatsAppEvent sign-ups traced back to the ad or form that brought them in.
An organisation that runs public events needed to know which promotions actually brought people in. Visitors sign up on the public site, and organisers see every sign-up in an admin dashboard traced back to its source, whether a Meta ad, a HubSpot form or a referral. It also drafts event posters and short promo videos with AI.
Next.js · React · TypeScript · Supabase · Vercel · Resend · HubSpot
Want something like this for your team?
Chat on WhatsAppWhat I can build for you, and the tools I build it with.
LLM tools that read, sort and draft for you, wired into the apps you already use.
Search and Q&A over your own files, with the source shown for every answer.
Scheduled jobs and bots that take repetitive admin off your team's plate.
Telegram bots with databases, admin pages, reminders and approvals.
Pull data from many sources on a schedule, cleaned and ready to use.
Reports that write themselves, in Power BI, Metabase or a web dashboard.
Prediction and scoring models, from property valuations to trading signals.
Next.js front ends, Python APIs, and servers set up with monitoring and backups.
AI comes standard in every package, not as an expensive extra. You pay for how much the system has to do. After a short call I send a fixed quote that settles where in the range your project lands, and nothing begins until you approve it.
One AI workflow, end to end
From my work: the receipts bot, the job application writer
Ask about StarterA connected system for one team
From my work: the Team Operations Suite, the event platform
Ask about BusinessOrganisation-wide or data-heavy
From my work: TradeIQ, the trading model, the property valuation models
Ask about CustomAI systems drift after launch: answers slip, costs creep up and models get retired. The care plan covers hosting, backups and monitoring, tracks your AI spend, fixes answers that drift and moves you to new models when old ones are retired. An hour of changes is included.
US$60 (S$80 · Rp1,1 juta) for work outside a quote.
Four years building data systems for companies in Singapore. Tested code, a test environment that mirrors production, and monitoring after launch are habits from that work.
BSc (Hons), National University of Singapore
Automated 8 quarterly reports and 4 indexes, cutting reporting time by more than 60%.
Moved the company onto a new Airflow setup, built client APIs and deployed machine-learning models on AWS Lambda.
Property valuation models with a 3% median error on Singapore HDB flats and condos.
Machine-learning classifiers and large-scale simulations.
Send a few lines: what you want automated, who will use it, and when you need it. I'll reply with questions or a fixed price.