Answers that always cite the source.
Find the answer in the corpus.
KilnAI is SunshineHouse's retrieval-augmented research platform for organizations with dense, technical corpora. Every answer cites a specific source. Every source carries an evidence tier. First in production at Atlantic Biocarbons; licensable for other research-heavy teams.
Fixed carbon in softwood biochar generally peaks between 500–650 °C1, though yields above ~600 °C trade off with total mass recovery2. Internal pyrolysis runs on Maine pine put the practical optimum near 575 °C3.
A serious answer over a serious corpus.
Biochar research lives in scattered PDFs, agency bulletins, manufacturer datasheets, conference proceedings, and internal lab notes. A real technical question usually has an answer somewhere in that pile. Finding it is a half-day.
Atlantic Biocarbons asked for a research surface their technical staff could actually use — one that turns the half-day into a question and an answer, without quietly hallucinating its way past the parts it doesn't know. KilnAI is the platform we built to do that. It does not invent. It does not pretend an internal lab note is a peer-reviewed result. It tells you what the corpus says, and what kind of source said it.
“If we can't show the source, we don't ship the answer.”
— the rule KilnAI is built around
Ask a real question. Get a cited answer.
A working researcher's loop, compressed. The question goes in. The corpus is searched by meaning. The answer comes back with its sources attached and graded.
Fixed carbon in softwood biochar generally peaks between 500–650 °C1, though yields above ~600 °C trade off with total mass recovery2. Internal pyrolysis runs on Maine pine put the practical optimum near 575 °C3.
Not every source is equal. We tag what it is.
Every document in the corpus is tagged at ingest by what kind of source it is. The model leans on the tags when it answers, and the tags surface in the citation. A reader can see immediately whether an answer is anchored in peer-reviewed work or in an internal observation.
Journal articles, conference proceedings.
Agency reports, technical bulletins, white papers.
Datasheets, spec sheets, marketing claims.
Lab notes, production runs, internal memos.
Built first-party. Built to be owned.
KilnAI is not a wrapper on someone else's chat product. The corpus, the embeddings, the typing, and the admin surface all belong to whoever is running the instance. If SunshineHouse walks away, the index stays.
Postgres + pgvector, owned by the licensee
The corpus and its embeddings live in a Postgres database the licensee owns. No vendor lock-in, no usage-priced black box. If SunshineHouse walks away tomorrow, the index is still there and still usable.
A real ingest pipeline, not a paste bucket
PDFs, scanned reports, internal lab notes, agency bulletins — each ingested with provenance intact, chunked sensibly, and re-indexable when the corpus changes. New documents are first-class, not bolted-on attachments.
Each source knows what kind of source it is
A peer-reviewed paper, a manufacturer datasheet, an internal observation, and a forum post are not the same thing. KilnAI tags every source by type at ingest, so the answer can weigh them honestly instead of pretending they're equal.
Answers carry a confidence the reader can see
Every answer is built from typed sources, so it can be graded — high-confidence when peer-reviewed agreement is clean, hedged when the evidence is internal or thin, unanswered when the corpus doesn't have it. No quiet hallucinations.
Every answer cites. Every citation links.
Citations are not a footnote feature — they are the contract. Every claim points back to the specific passage in the source document the model leaned on, one click away. A researcher can verify the answer, not just trust it.
Cheap to run, easy to keep
Standard Next.js on Vercel, a managed Postgres instance, and an admin surface a subject-matter expert can use without an engineer in the room. No infrastructure heroics required to keep it standing.
Four steps. No black box.
Ask
A real question in plain language — the kind a researcher would type into the margin of a paper.
Retrieve
The corpus is searched by meaning, not keyword. Every candidate passage carries its source and source type.
Grade
The retrieved evidence is weighed by type — peer-reviewed beats manufacturer claim beats anonymous forum post.
Cite
The answer arrives with inline citations. Click any one to read the passage it leaned on, in context.
A research surface for any technical corpus.
KilnAI is a SunshineHouse product. Atlantic Biocarbons is the first organization to run a licensed instance, loaded with their own biochar research. The platform is built to be loaded with other corpora too — conservation science, soil and watershed data, regulatory and grant literature, a research lab's own back catalog.
Each licensee gets their own instance, their own database, and their own corpus. Source typings can be tuned to the field — what counts as “peer-reviewed” or “grey literature” looks different in soil science than in clinical research, and the system is set up to reflect that.
Atlantic Biocarbons
A licensed instance, loaded with ABC's own biochar research corpus.
Try the public preview ↗A research corpus near you
Conservation districts, soil-science labs, grant-funded research programs, regulatory teams — any group that already has a dense technical corpus and not enough hours to read it.
Talk to us about licensing →If we can't show the source, we don't ship the answer.
The rule sounds obvious. Most chat products built on the same underlying models do not follow it. KilnAI does, by default — citations are wired into the answer pipeline, not added at the end.
The point is a research tool a working scientist can use without having to second-guess it. Anything less is a slot machine for footnotes.
Have a corpus that needs reading?
KilnAI is live, in production, and ready to load with the next corpus. We're talking with research-heavy teams about what an instance would look like for their field.