Projects

Applied AI systems, research, and the infrastructure underneath them.

Applied AI Products

FiskalLink: an analysis that took five expert-days now takes twenty minutes

A ministry-funded system that broke a schema incompatibility standing for ten years. The reconciliation an expert used to perform by hand, over a working week, is now a rule applied once and an analysis anyone can run before lunch.

Klaussa AI: building a legal AI company on a small team

Founding engineer on a vertical AI platform covering the whole Indonesian regulatory corpus. Agentic legal research, smart search over regulation, hierarchy visualisation, opinion drafting and court simulation, with the infrastructure chosen so a small team could run all of it.

KUHPedia: a chatbot over the criminal code, used by about 2,000 people

A conversational interface to the Indonesian criminal code, built inside a ministry partnership. Roughly 2,000 users, which makes it the piece of my work that has been used by the most people.

Harmonisator: finding the places where regulation contradicts itself

A tool built with the Ministry of Law to detect overlap and conflict between Indonesian regulations, where the volume of instruments makes manual review impractical.

CariKerja: measuring a market until the product had to change shape

An AI job-application agent for Indonesian job seekers, built Cloudflare-native. Four coverage measurements taken before writing the expensive part showed that the auto-apply premise borrowed from the US market does not transfer here, and the product was rewritten around that finding rather than against it.

ResepCerdas: screening a prescription photo without letting the model decide

An Indonesian prescription screening aid that reads a photo of a paper prescription and runs it through 34 pharmacy checks. The language model transcribes; every verdict is deterministic code, and the checks that cannot run say so instead of reporting green.

Research

Klaussa-8B: replacing four API calls with one model

A Llama-3.1-8B specialization trained to do in one pass what four chained Gemini Flash calls did before, taking the projected cost of answering a legal question from USD 0.09 to USD 0.0012. Written from the middle of the work rather than after it.

Long document summarization: a 3B model that beats a 70B

Summarising documents of up to a million tokens by reframing the task as question answering. A 3B model answering 24 learned questions by retrieval produces better summaries than a 70B model reading the whole text, because the standard framing of summarisation is underconstrained.

Steering emotion in a speech model without training it

Testing whether a text-to-speech model already contains prosody controls nobody exposed. Treating inference-time control as an optimisation problem over latent dimensions, with an emotion encoder supplying the objective.

SemEval-2025 Task 11: the purpose-built encoders lost

First-author system paper on multi-label emotion detection. A wide sweep across state-of-the-art sentence encoders and classical classifiers, against large prompt-based models. The specialised encoders did not win, which was not the result we expected.

Prompt compression at EleutherAI: rebuilding Selective Context on a harder benchmark

Four weeks of remote research with mentors from EleutherAI and Apart Research, re-implementing a published prompt compression method on benchmarks it was never tested against. The contribution is the reproduction itself: moving a 2023 technique onto LongBench v2 and MS MARCO, and recasting it as a token level classification problem.

Consumer health QA: retrieval that fires when the model hesitates

A retrieval augmented question answering system for Indonesian consumer health questions, built so that it goes back for more evidence mid sentence whenever the model produces a low confidence token. First author work, three person team, second place at Gemastik 2024 Data Mining.

Beating a proprietary embedding API with a 568M model on Indonesian legal search

A domain-specific retrieval encoder for Indonesian regulation, trained on one consumer GPU. It significantly beats the commercial embedding API it was built to replace on the query distribution the product actually serves, and ties it on the other, at roughly an order of magnitude fewer parameters. The result rests on two benchmarks built before any model was trained.

Infrastructure and Tooling

Reading 300,000 government documents for 8% of the going rate

Scraping and OCR infrastructure for the Indonesian regulatory corpus. Self-hosted document layout and vision models on owned GPUs cut the OCR budget by 92% across more than 300,000 documents.

Miss Scraper: a browsing agent in 1,257 lines of Python

An open source web scraping agent that drives a real browser from a plain language instruction and returns validated JSON. Built as a deliberately small alternative to browser-use, with the browser exposed to the model through seven tools over the Model Context Protocol.

Indexing 170,000 AI conference papers: what got built and where it stopped

A pipeline that scrapes metadata for every paper from fourteen top computer science conferences, extracts the text from each PDF, and stores it as markdown in object storage. The metadata index is complete at 169,864 papers. The text extraction reached 1.5 percent and has not moved since March 2026.

Gapfinder: the model proposes research gaps, the corpus decides

A pipeline that reads a local corpus of AI papers, proposes candidate research gaps in a topic, and judges whether each is solvable today with low effort. Every claim is grounded in retrieved evidence rather than model opinion, and the system reports its own degradation instead of hiding it.

Paperbase: a local index of 164,000 AI papers built for agents to search

A self-updating index of the AI research literature, from conference rosters through metadata enrichment to extracted full text, exposed to language model agents as a small set of query tools. Built so an agent can check whether a problem has already been solved without depending on a hosted search API.

Efficient life through proactive AI: an assistant that speaks first

Every assistant I run waits to be asked, which means it only helps when I remember that it could. This is the design work on the opposite arrangement, a cheap model watching continuously and deciding when something is worth escalating to an expensive one.