PortfolioCase study
RC-01Live
RailCite
Research on track.
A trust-first assistant for Indian Railways commercial circulars: it cites the right circular with number, date and supersession lineage — or refuses.
5,760
documents indexed
Measuredas of 15 Sep 2026
live corpus; a nightly crawl keeps adding circulars
0
invented citations
Structural
by construction: the validator drops any citation that doesn't resolve
68%
of ingested PDFs needed OCR
Measuredas of 7 Sep 2026
3,865 of 5,687 ingested PDFs
Evidence keyMeasuredStructural
The problem
One wrong circular can damage an inspector’s credibility.
A Chief Commercial Inspector has to defend every demurrage or wharfage decision — across years of circulars, scanned PDFs and rules that were quietly superseded.
- Yearly PDF lists
- Scanned circulars
- Guess the current version
- Check supersession
- Write the justification
- Cite by hand
Instructions not included “should not be deemed to have been superseded simply because of their non-inclusion.”
The product
Ask, get the right rule — or a clear refusal.
The inspector asks in plain language. RailCite answers only from passages that govern the case, and says so when none does.
![RailCite's Ask console: a demurrage-waiver case in the Commercial domain, answered “Cited from 8 passages in the Goods manual” with an inline citation marker [1] and an English/Hindi toggle.](/media/case-studies/railcite/answer.webp)
Answered
- The governing circular, with number and date
- Supersession lineage: which version governs today
- Every citation resolves to a real source
Refused
- No passage governs the case
- Designed as a success state, never an error
Product decisions
Refuse rather than fabricate.
- Could have
- Always answer with the best-matching passage
- Chose
- Cite-or-refuse: refusal is a first-class success
- Because
- A confident wrong citation damages the inspector’s credibility, not the tool’s.
- Could have
- Trust the prompt to behave
- Chose
- Validate every citation in code
- Because
- A validator drops any answer block whose citations don’t resolve; if every block drops, the answer becomes a refusal.
- Could have
- Make cross-domain bleed unlikely
- Chose
- Make it impossible with a hard domain filter
- Because
- One commodity’s circular must never reach another’s case: “bleed has to be impossible, not merely unlikely.”
How trust works
Trust lives in the system, not just the prompt.
- Questiona signed-in request
- Embed + classifyVoyage-3 vectors; Claude Haiku routes the domain
- Retrievetop 8 passages, hard domain filter
- Thresholdcalibrated 0.45 → 0.32
- SynthesisClaude Sonnet 5, extractive, answered | refused
- Citation validatordrops citations that don’t resolve
- Answer or refusalwith supersession lineage
Evidence
What the records show — and what they don’t.
The retrieval threshold was calibrated against real queries rather than guessed, and the corpus runs live. Demand-side proof doesn’t exist yet.
0.32
calibrated relevance threshold
Measured
5 relevant + 3 irrelevant queries: irrelevant ≤ 0.25, relevant 0.29–0.66; a nonsense query was refused
14,406
chunks indexed, live
Measuredas of 15 Sep 2026
193
supersession lineage links
Measuredas of 7 Sep 2026
22/40
UX critique of the Ask screen
Measuredas of 31 Aug 2026
one P0, “Flagship starter refuses”; no fix is recorded
What isn’t measured
- No usage data and no measured time-to-cited-answer.
- No groundedness, retrieval-precision or latency eval; citation validity is structural, not sampled.
- Sessions with real inspectors happened but weren’t logged.
Key learnings
Three lessons that made RailCite trustworthy.
Refusal is a feature
A confident wrong citation is worse than no answer.
Trust belongs in code
Citation validation and supersession checks enforce correctness beyond prompting.
Freshness is correctness
A superseding circular turns a perfectly cited answer wrong — so the corpus is crawled nightly.
Sources
Everything here is backed by a real artifact.
- PRD
- Design
- Architecture
- Code
- Evaluation
- Test run
- Live data
- Build ledger