ReguLens

• 618

  Automated tests

  Green from a clean checkout with no GCP account and no cost

• 25.5s

  Upload to alert, single pasted rule

  Measured on the deployed stack

• 174.3s

  Upload to alert, 55-clause EU annex extract

  Extraction accounts for roughly 72% of it, bound by output tokens

• 6/6

  Country discovery accuracy

  Regulator names and domain roots both correct six out of six; every model-written path was wrong, which is why paths are read from fetched pages

• 88

  Self-discovered regulation

  Verbatim limits read out of Commission Regulation (EU) 2023/2108, which the system found in CELLAR unprompted

1. 01

   Worked across both sides of the stack (the FastAPI backend and the Next.js operations console) as one of two contributors, authoring 39 of 83 commits

2. 02

   Built and corrected the impact engine so a verdict names the rule it actually rests on and a requirement never outlives the clause it came from

3. 03

   Made alerts name and link the regulation that moved a verdict

4. 04

   Extended the guardrail to read food category, letting the review queue answer more cases without a human

5. 05

   Fixed query retrieval to target the market a question names rather than only its wording, and to stop treating a country as a jurisdiction

6. 06

   Reworked extraction to flush each PDF page instead of holding an entire document in memory

7. 07

   Laid out the web app as an operations console rather than a set of pages

1. 01

   Compliance twin: a product with ingredients and quantities, targeted at one to three destination markets

2. 02

   Ingestion by PDF upload, pasted text, or a daily scheduled re-read of every watched address, no third path

3. 03

   Verbatim clause extraction with a composite confidence score, never a summary

4. 04

   Deterministic guardrail deciding whether two clauses may be compared at all

5. 05

   Reconciliation verdicts of created, superseded, conflict opened, or flagged for review, exactly one event per clause

6. 06

   Impact engine as pure arithmetic, with no agent and no model call

7. 07

   Immutable audit event written in the same batch as every state change

8. 08

   Alerts naming the regulation that moved a verdict and whether anyone uploaded it

9. 09

   Four kinds of watched source (document, feed, listing, and SPARQL catalogue) because a change means four different things

10. 10

    Country discovery that produces an index address to watch, then stops

11. 11

    Query agent that picks its own retrieval tools and refuses to answer without citing a stored clause

12. 12

    Readiness reported as a count of issues rather than a percentage

One container image across four Cloud Run services, with everything slow placed behind Pub/Sub: the API hashes, stores, publishes, and returns 202. Three topics carry the pipeline: document uploaded triggers extraction, clause extracted triggers reconciliation, graph changed triggers impact, and any topic dead-letters after maximum retries. Deterministic code owns every mutation; a model response never reaches clauses, requirements, or conflicts without passing a Pydantic validator and the guardrail. The core package imports neither FastAPI nor ADK, so every tool body is an ordinary function testable without a framework.

Front end

Next.js, React, TypeScript, Tailwind CSS

Back end

Python, FastAPI, Google ADK, Pydantic

Data

Firestore, Cloud Storage

Infrastructure

Google Cloud Run, Pub/Sub, Cloud Scheduler, Cloud Build, Artifact Registry, Secret Manager

Tools

Docker, Make

Integrations

Gemini 3.5 Flash, Gemma, Vertex AI, EUR-Lex CELLAR, BPOM

1. 01

   Problem

   A system that confidently declares a false conflict destroys the trust the whole product rests on, and an LLM judge asked to compare two clauses will always produce an answer.

   Solution

   A deterministic guardrail decides whether two clauses may be compared at all (substance family, product type, unit, jurisdiction) before any model is consulted. Different jurisdictions with different limits resolve to a conflict in code. The LLM judge runs only on same-jurisdiction pairs whose effective dates leave the supersede question open.

2. 02

   Problem

   Country discovery needed regulator URLs, but the model's URL paths were wrong every single time; measured at six regulator names correct out of six and six domain roots correct out of six, while every path it wrote was wrong.

   Solution

   Asked the model only what it can actually do (name the regulator and its domain root) then read paths from pages actually fetched. The model picks an index from a supplied link inventory, and any choice outside that list is discarded.

3. 03

   Problem

   Content hashing on raw bytes never matched, because EUR-Lex stamps a fresh session id into every response; the same document looked new on every fetch.

   Solution

   Hashed the extracted text rather than the bytes, so the cache short-circuits genuinely identical documents that have already reached a terminal state.

4. 04

   Problem

   The hackathon required Gemini 3.5 or newer, but the asia-southeast1 region only offered gemini-2.5-flash, using the nearest region would have broken a hard entry rule.

   Solution

   Split the topology: infrastructure stays in asia-southeast1 while Gemini is reached through the global endpoint, with multilingual embeddings alongside it.

5. 05

   Problem

   Pub/Sub delivers at least once and will redeliver, so a naive pipeline produces duplicate clauses and duplicate alerts for a single document.

   Solution

   Made every handler idempotent through a state check, with redelivery tested explicitly rather than assumed, and Firestore transactions around each clause mutation so parallel reconciliation cannot race on a shared clause.

6. 06

   Problem

   A requirement generated from a clause kept applying after that clause had been superseded, so a product could fail against a rule that no longer existed.

   Solution

   Retired requirements whose source clause was superseded at every materialization, and made each alert name the rule its verdict actually rests on rather than the one that started the run.

1. 01

   An honest needs_review beats a confident wrong answer. Low confidence or low authority queues a rule for a human and never quietly shifts a limit.

2. 02

   Readiness is shown as a count of issues rather than a percentage, because a percentage needs a denominator nobody actually has.

3. 03

   Being precise about which parts are genuinely agentic mattered more than the number of agents; one of the four makes no model calls at all.

4. 04

   A flag that runs the entire pipeline with no model calls kept the local stack and most of the test suite free to run.