FlatFeed
Product case study
01Problem
Saved-search alerts on real-estate portals only cover listings published on their own platforms. Meanwhile, housing providers are a primary source of WBS listings and may publish them on their own websites before they reach real-estate portals. Some providers offer only email alerts; others offer none. Users therefore repeat the same search across several places and compare different listing formats. FlatFeed tests a simpler flow: save your criteria once and review matching listings in Telegram.
02Product
Working Telegram prototype · Generated test listings · Rule-based matching · Bounded AI QA evaluation
Set WBS type, district, maximum Base rent, excluding operating and heating costs. and rooms once.
Each permitted source would map into the same listing format. Structured feeds can supply fields directly; text-based or incomplete sources require extraction. This prototype implements the text-based path.
Code compares each listing with the saved criteria. If a required value is missing, the listing does not match.
Users can request matches when they want. FlatFeed returns each match in a Telegram card. After a user saves a filter, FlatFeed automatically sends each new match in Telegram and records each successful send to prevent duplicates.
FlatFeed opens with one clear action: set up a filter.
Select the certificate type used for matching.
Select one Berlin district or keep the search open.
Set the highest base rent that should match.
Complete the four-field filter.
Review or edit the criteria before checking listings.
See eligibility, rent, apartment details and estimated walks to S- and U-Bahn stations in one card.
Photo: Bodo Kubrak · CC0 1.0
Apartment details are synthetic.
The prototype currently uses generated listing text rather than live provider data. It implements one possible ingestion path: deterministic extraction into a common listing format with an admin-only AI quality check.
How listings reach FlatFeed in production depends on the access method, data format and information each provider supplies. Complete structured feeds can be added directly; text-based or incomplete sources may still need extraction and review.
Automatic delivery and the runtime AI check are built but disabled in this prototype. The AI-check results come from an offline evaluation using synthetic listing data, not listing data from real providers.
03Decisions
A
For this self-directed portfolio project, I deliberately tested one plausible text-based ingestion path to demonstrate how I scope, evaluate and constrain AI. The model checks parser evidence but never changes listings or decides matches. If a provider supplies complete, reliable structured fields, its adapter should map them directly; the parser and AI check are used only where extraction is actually needed.
B
I chose generated listings to test filter setup, normalization, matching and Telegram delivery without using provider data until source access and reuse terms are clear. This tests product mechanics, but not source coverage, listing freshness or renter value; those require permitted live sources and a real-user pilot.
C
If rent or room count required by a filter is unknown, the listing does not match. This reduces unsupported matches but may hide a suitable apartment, so the trade-off still needs real-world validation.
04AI Evaluation
This evaluation covers the prototype’s text-based ingestion path. AI independently checks the fields extracted by the parser and flags differences for admin review; rules still decide every match. A provider feed with complete, reliable structured fields could bypass this path.
Illustrative planted-error example.
Listing text
Charlottenburg-Wilmersdorf · 2 Zimmer · 512,40 € · WBS 100–140.
Parser output
AI evidence
Mismatch detected. The parser missed WBS 100 and could hide a suitable listing. Admin review required. AI flags the difference and stops there. A configured admin compares the original listing and marks the finding as parser error, parser correct or unsure. Nothing changes on the model's word alone.
Before testing any model setups, I defined the metric targets. I then compared model setups on separate development data. GPT-5.6-Terra with high reasoning met those targets, so I froze the setup before running it once on the locked 600-listing evaluation. FlatFeed already includes an admin-only AI check that asks a model for source evidence and compares it with parser output. To test one model configuration for that check, I ran it once on 600 synthetic listing pairs: 300 clean and 300 with one planted parser error. For each field, the model returned an exact source quote or no value, and code checked it against the parser output. There were no retries or tuning after the run.
Prototype target: at least 98% (294/300). Target met.
Prototype target: at most 1% (3/300). Target met.
Prototype target: at least 98% (294/300). Target met.
Prototype target: at least 99.5% (597/600). Target met.
An overall score can hide a weak spot. This breakdown shows how the check performed on each listing field.
| Listing data | Used in | Errors found | Target | Result |
|---|---|---|---|---|
| WBS | Matching + Telegram card | 75/75 · 100.0% | At least 74/75 | Target met |
| District | Matching + Telegram card | 30/30 · 100.0% | 30/30 | Target met |
| Kaltmiete | Matching + Telegram card | 60/60 · 100.0% | At least 59/60 | Target met |
| Rooms | Matching + Telegram card | 50/50 · 100.0% | At least 49/50 | Target met |
| Address / postal code | Telegram card | 40/40 · 100.0% | At least 39/40 | Target met |
| Floor | Telegram card | 25/25 · 100.0% | 25/25 | Target met |
| Warmmiete | Telegram card | 20/20 · 100.0% | 20/20 | Target met |
Berlin reported 12,398 apartment re-lettings by six state-owned housing companies in 2024, excluding Berlinovo. Because no public source gives a combined count of unique listings, I used 15,000 checks per year as a rounded planning scenario. View the Berlin source.
Based on the measured run
About $35/year $2.94/month · 15,000 × $0.00235 per checkThe final 600-check run cost $1.412906.
Planning case · 25% buffer
About $45/year $3.68/month · 15,000 × $0.00294 per checkAdds a simple buffer for variation in listing length and model output.
This estimate covers AI API calls only. It assumes one check per unique listing for each prompt version and uses prices recorded on 30 July 2026. It excludes source access, hosting, monitoring, duplicate listings, price changes, differences in real-listing length and human review. The 12,398 re-lettings show the possible scale, not the number of ads a live system would process.
05Next
I built and tested the saved-filter Telegram flow and evaluated one plausible way to process text-based listings with bounded AI QA. Matching remains rule-based; how listings enter FlatFeed in production should depend on the access terms, formats and data quality providers actually offer.