An AI-assisted institutional archive for Simhastha Kumbh Mela 2027. Every letter, minute and order, from receipt to cited answer.
1,027
documents indexed
28
departments covered
6,701
pages OCR'd
2015–2027
correspondence span
< 1 min
to a cited answer
The problem
Operational needs and challenges
Simhastha is coordinated through correspondence. Every commitment, approval and instruction sits in a letter or a minute, and has to be found, read and acted on correctly. An archive that serves officials needs six things, each with a known obstacle.
Operational need
What gets in the way
How the Tracker answers
1Find it fast
Thousands of pages across inboxes, files and folders; nobody remembers the file name or date
Ask in plain language; hybrid search finds the letter by meaning or by exact number
2Read every page
Scanned letters can't be searched; Marathi and English mixed in one file
Every page OCR'd; Marathi translated; the original page kept beside the text
3Trust the answer
A summary without its source can't be quoted in a meeting or a reply
Every answer cites its letters; the source PDF opens beside it
4Catch commitments
Deadlines and instructions sit in paragraph four of a long letter
Action items extracted with owner and date, tagged by criticality
5See the whole picture
Each department holds only its side of a conversation
Entities linked across letters; wiki pages and a graph of who wrote to whom
6Keep the memory
Transfers take the context of past decisions with the officer
Every document, summary and wiki page kept on record, with history
COST OF A MISS :A letter not found means a decision re-made, a deadline missed and delay before the Mela.
What the Tracker is
One flow from letter to answer
The Archive Tracker is an AI-assisted institutional archive: it takes every inter-department letter from upload to a cited answer, and officials verify against the source.
Nothing is entered twice: the text read at upload drives search, action items, the wiki and the graph, and every answer points back to the original.
Stage
What happens
1Upload
Letters, MoMs, notices and orders added, scanned or digital, from any departmentKept on record: the original PDF and its date.
2Read
Each page OCR'd; Marathi translated into English; full text kept page by pageKept on record: page text and translation.
3Understand
Summary, document type, sender, departments, people, amounts and dates extractedKept on record: document profile and entities.
4Index
Text split into passages and embedded for meaning-based plus keyword searchKept on record: searchable passages.
5Act
Instructions and deadlines become action items, auto-tagged by criticalityKept on record: action items and alerts.
6Connect
Topic, entity and analytics wiki pages compiled; correspondence graph updatedKept on record: wiki pages with change history.
7Answer
Officials ask, read the cited answer, open the source and draft a replyKept on record: the reply letter, saved back as PDF.
See it in action
The portal, screen by screen
Real screens from the live archive. Click any screen to view it at full resolution.
01
Library
Every letter in one searchable register
1,027 documents with date, department, direction, type and parties, filterable in a click.
Filters by direction, type, department and uploader, with live counts
Insight badges flag budgets, action items and linked letters per row
Export the filtered view to CSV
02
Document view
The original beside what AI read from it
Open any letter to see the source PDF next to its extracted profile; here, a Marathi order releasing ₹1,200 crore to NTKMA.
Sender, recipient, date and type read from the letter
AI summary in English, even for Marathi originals
Budget lines and directives pulled out with owner and deadline
03
Ask AI
A plain-language question, a cited answer
Officials ask the archive directly. The answer quotes the order number, amount and condition, and lists the letters it drew on.
Inline citations to the source documents
Brief, Normal or Detailed answers, or an exact word count
Answer only from filtered documents; save an answer as a knowledge page
04
Dashboard
The whole Mela at a glance, for senior review
Documents, OCR'd pages, open action items and budgets across 28 departments, without opening a single file.
₹686 Cr of proposals tracked by department, estimated vs approved
3,034 action items by status, with the top open items and owners
Updated as each new letter is added
05
Archive
Institutional memory, written from the letters
An auto-compiled knowledge base of documents, entities, locations, financials, tasks and topics, linked to the source correspondence.
16 topic pages, from Ghats & Riverfront to Police & Crowd Management
Entity pages ranked by correspondence volume
Edit, rebuild and download any page as PDF
Architecture
How the Tracker is built: layers, controls and considerations
People
NTKMA officials and staff; every decision and reply is theirs
↕
Portal
Dashboard, Library, Ask AI and Wiki in one signed-in web app
↕
Controls A wrong answer is caught before anyone relies on it
1Grounded retrievalThe model answers only from the passages retrieved
2Cited sourcesEach answer lists the letters it drew on
3Original keptThe source PDF and page text are never overwritten
4Human reviewOfficials read the source before acting or replying
5Fails safelyNo passage found means no answer is invented
↕
Model layer
Local Qwen3 on the server for most tasks; Gemini where a task needs it
↕
Data and background engine
Postgres with pgvector, file storage, page-level OCR, translation and indexing
Design considerations
Self-hosted
Archive, database and search run on one dedicated server, behind sign-in.
Local model first
Qwen3 runs on the server itself; a cloud model is used only where a task needs it.
Search quality
bge-m3 embeddings with Postgres pgvector, combined with keyword matching.
Bilingual by design
Marathi and English pages are both read, translated and searchable.
Recoverable
Database backups and documented restore; wiki edits can be reverted.
Why the Tracker
Key unique selling points: what sets it apart
What did PWD commit to on the Ring Road Phase 2 alignment?
PWD confirmed the revised alignment and asked for land handover [1]; the joint meeting minuted the timeline and assigned follow-up [2].
[1] PWD letter to NTKMAOpen ›
[2] Minutes of joint review meetingOpen ›
Illustrative answer view: the question, a written answer, and the letters it was drawn from, each one click away.
Examples only. Questions officials ask every day, such as: Commitments: what a department agreed to, and by when Approvals: which meeting approved a budget or design Follow-ups: open action items for a project or officer Replies: a draft response to a letter just received
Source beside every answer
Answers cite the letters they came from; the PDF opens alongside, so a claim is checked in seconds
Ask, don't search
No file name, date or sender needed; meaning-based and keyword search run together
Built for Marathi and scans
Scanned pages OCR'd and Marathi translated, so no letter is left out by format or language
Commitments never buried
Action items pulled from every letter with owner and deadline; critical ones raised on the dashboard
Connected across departments
People, projects and proposals linked; wiki pages and a graph trace a decision end to end
Chat inside a document
Ask about one open letter, or select a passage and ask about just that
Reply-ready and on record
Draft a response from the original's context; everything kept for the Mela and after
In summary
Four things to take away
Findable
Any letter, minute or order in under a minute, asked in plain language.
Traceable
Every answer cites its source, which opens beside it for checking.
Complete
Scanned and Marathi pages read, so nothing is missed by format or language.
Lasting
Institutional memory that stays when officers move on.
The Tracker finds the record.Officials make the decision.
