Internal
Tasalee, working end

Everything the client does not see.

The public site sells one number. This is where the number is made: the resolver, the government fetches, the scoring engine, the agent that learns how Shubham rates, and an honest register of what is still provisional. Nothing here is client facing.

Analyse a property › Train the agent

Dashboard

Resolve a coordinate to a parcel identity, queue the sources that branch calls for, score the property and pull the PDF.

Open ›

Reports

Both sample health cards, the free screen and the full premium file, with the findings and the price argument behind each.

Open ›

Data layer

Nineteen sources with what each proves and what it does not, the three branch resolver, and the fetch tools.

Open ›

Agent

Feed it a file you have already scored by hand. It fits weights to your judgement and reports its own error before it is trusted.

Open ›

Assistant

Ask it about khewat, intaqal, a mutation code, a unit conversion or a parameter rubric. Answers come from the mapped ontology, not a guess.

Open ›

Mail desk

Every website enquiry arrives at contact.tasalee@gmail.com. Read it here, let the agent draft the reply from what is actually mapped, and answer from that inbox.

Open ›

Where the build actually stands

Written plainly so nothing gets sold before it works.

Three branch parcel resolver Revenue estate, licensed colony, condominium. Most specific identity wins and the superseded layers are still reported.
Running
Jamabandi, mutation, deed and collector rate fetches Cascading dropdown session held, viewstate preserved across postbacks, per host throttle.
Running, captcha gated
DORIS, eCourts, CERSAI, RERA, DTCP, air quality Six secondary readers behind the same session layer.
Running
Scoring engine, four asset classes Two hundred fifteen parameters, category weights per class, thirteen findings that cap rather than deduct.
Running
Rater agent Non negative ridge fit on hand scored files, leave one out error reported every time it scores.
Running, two seeded files
PDF health card Designed header band, running footer, logo derived watermark on every page, full parameter appendix.
Running
Cadastral geometry Twenty pilot footprints are provisional squares around approximate centroids. They place a pin in the right pocket, never on the right plot.
Provisional
HSVP and DTCP layout digitisation Real plot numbers need the sector layout plans traced sheet by sheet. Not started.
Not built
Captcha solving Deliberately not automated. The image is handed to an operator, answered, and the session continues.
By design

How much of a score the record can carry on its own

Weighted by category, per asset class. The rest needs a site visit, a counter, or a document the seller has to produce.

Eleven of the thirteen capping findings can be tested automatically, which is the number that matters more than the coverage percentage. A file can be cheap to screen and still be caught by the one thing that sinks it.

Standing rules for anything that leaves this console

State the coverage beside the score

A seventy on eighty percent coverage and a seventy on forty percent coverage are different claims. Both go on the page.

Name the source and the date

Every finding carries the source it came from, its evidence tier and when it was last refreshed. An assertion by a seller is tier six and is labelled as one.

Never publish the provisional geometry

Pilot footprints go no further than this console. A client is told the pocket, and told that the plot boundary is unverified.

The agent does not score alone

Below three hand scored files it keeps framework weights and says so. Above that it reports its leave one out error next to every number it produces.

Disclose what a source cannot prove

A jamabandi is a record of rights, not a title guarantee. A RERA registration is not a construction warranty. Both are written into the source map.

No client sees this vocabulary

Knockout, tier, divergence, viewstate and coverage are internal words. The client side of the same fact is written in plain language.

The engine, in numbers

Parameters
215
scored across four asset classes
Sources
19
nine open, six captcha, two login, one counter, one RTI
Capping findings
13
eleven testable without a site visit
Backend routes
27
resolve, records, score, train, card, tools
Mutation codes
12
mapped Devanagari to instrument type
Assistant entries
102
across nine topics, inverted index

Open the funnel

Drop a coordinate, watch the resolver pick a branch, and run it through to a verdict.

Open the dashboard ›

Pin, parcel, record, score.

The seven step funnel from the deck, running end to end. Resolve a coordinate to a parcel identity, see which government sources that branch calls for, then score the property on the engine the paid report uses.

Sample data

Running on built in sample data. Start the backend and paste its address to fetch live records, generate PDFs, and train the agent on your own scoring.

Property shortlisted
Parcel resolved
Records queued
Parameters scored
Verdict and price

Locate

Assess

Locate by anything

Parcel identity

Open plot
HEALTH SCORE
45/ 100
High caution

BEFORE ANY KNOCKOUT
74.5/ 100

COVERAGE
12/ 12

Records this branch calls for

Open sources run on their own. Captcha gated ones queue for a person, which is deliberate.

Parameters

Each row scores one to five against its own rubric, with the evidence tier behind it. Hover a name to read the rubric. Rows marked KO cap the score.

ParameterCategory ScoreEvidenceWeightState

What a client actually receives.

Both formats take the records section from a professional title search and the summary layout from a consumer inspection report. One number, the verdict in a few words, the rupee figure, then the problems in plain language before any parameter table.

Free title and location screen
Open plot, Pargati Nagar
Shastri Nagar dairy belt, Sonipat, Haryana   500 gaj   Owner of record Mukhtiyari Devi   Report TSL 2026 001
HEALTH SCORE
45/ 100
High caution

Material defects. Do not pay until they are resolved.

BEFORE THE KNOCKOUT
74.5/ 100

Capped at 45 by the missing land use permission.

NOC from relevant authorities
COVERAGE
12/ 12

applicable parameters assessed, mean evidence confidence 0.71

What is wrong here

Residential NOC is not on recordKnockout

Land use permission is pending. Until it is resolved the plot may not be buildable, which is not a deduction, it is a different asset. No payment should be made against this plot until the permission is produced.

Liquidity is lowMaterial

A dairy farm micro market with railway proximity. The buyer pool is small and time on market for comparable plots runs beyond twelve months.

Asking price is 16 percent above the micro marketNegotiable

Comparable registered transactions in the same revenue estate settle between ₹15,000 and ₹15,500 per gaj across the last four quarters. The ask is ₹18,000.

Title chain is consistentVerified

Owner of record matches the seller across the jamabandi, the mutation register and the registered deed. No link rests on a power of attorney and no litigation was found against any name in the chain.

Parameters assessed

ParameterScoreEvidenceState

Coverage and limits

Revenue record as at the 2016 to 2017 jamabandi for the village, tehsil Sonipat. Online copy, unsigned and marked reference only. HARIS deed records before 2004 are not digitised for this sub registrar and were not searched. Four of the twelve parameters rest on field verification rather than a government record, which is why the mean evidence confidence is 0.71 rather than higher.

Not included in the free screen. Negotiation price band, structural assessment, five year projection, the proof documents themselves, and the remaining parameters for this asset class.

A coordinate does not give you a khasra.

Across most of Gurugram the khasra was extinguished when the land was acquired and relaid, or it never applied at unit level at all. So the resolver has three branches behind one input, and what it promises is parcel identity rather than a khasra number.

19
sources mapped
11
of 13 knockouts automatable
44
parameters fed automatically
20
pilot footprints digitised

Gurugram resale stock is three different things

Revenue estate land

Wazirabad, Nathupur, Sikanderpur Ghosi, Chakkarpur, Kanhai, Badshahpur and the Sohna belt. Khasra and khewat are live and the jamabandi is authoritative. This is the only branch where a coordinate genuinely gives you a khasra, once the cadastral sheet has been rubber sheeted onto real coordinates.

HSVP sectors and licensed colonies

DLF Phases 1 to 5, Sushant Lok, South City, the numbered sectors. Once land is acquired and relaid the khasra stops being the property's identity. Identity is sector, block and plot, and the records sit with HSVP and DTCP rather than the patwari. The DTCP licence schedule lists the khasras the colony was carved from, which is the only bridge back to branch A.

Apartments and condominiums

Camellias, Magnolias and every tower like them. There is no parcel at unit level at all. The land is held as an undivided share and the unit is a floor space right, so identity is project, tower and unit, proved by the conveyance deed and the RERA registration.

Most specific identity wins

A single pin can sit inside a village, a colony carved out of that village, and a tower inside that colony. Specificity runs C, then B, then A. The superseded layers are still reported, because the revenue estate underneath is where the licence schedule bridges back to a khasra.

Nineteen places a fact can come from

Each one records what it proves, what it does not prove, how stale a cached answer may be, and the disclosure line the report has to print.

SourceProvesDoes not prove AccessTierRefresh

What is automated and what is not

Portals are discovered and monitored automatically. Gates that exist to stop machines, captchas above all, are cleared by a person. The artefact that carries legal weight is the certified copy ordered through the official channel, so nothing here substitutes a scrape for a certified record. The automation exists for speed, for surveillance, and for telling an analyst exactly where to look.

Look at the parcel before you trust the record

Paste a plus code, a coordinate, or a Google Maps link. Every imagery source that covers Gurugram and Delhi opens on the same point, so you can compare what the record claims against what is actually standing there. It opens on our own address until you give it something else.

Waiting for a point

A plus code off a Google Maps card, such as F38V+W3G, Sector 27, Gurugram, resolves here without a network call. A shortened maps.app.goo.gl link has to be opened once so it expands, then paste the long address from the browser bar. Nothing leaves this machine, the buttons only build addresses.

Satellite imagery shows what is built. It does not show who owns it, what is sanctioned, or where the boundary legally runs. Use it to catch the obvious contradiction, never as evidence in a report.

How much of the score the machine can reach

Forty four of two hundred scored parameters have an automated source behind them, which sounds thin until you weight it. The automated ones are the heavy ones.

The thirteen knockouts

Eleven can be checked without leaving the desk. The two that cannot are physical, which is the correct answer rather than a gap.

KnockoutWhat it meansCeilingReached by

Nothing joins to anything

The jamabandi is keyed on khewat and khasra. The deed registry is keyed on party name and date. There is no shared identifier anywhere, so three heuristics do the work. All three are running below.

Owner name matcher

Each name is reduced to a consonant skeleton that survives transliteration, then compared token by token, with parentage weighted equally because it disambiguates far better than a given name.

Khasra lineage

Compound khasra numbers read right to left and record a history of subdivision. A deed from 1998 citing khasra 155 still describes the parcel a 2024 record calls 5073 slash 4829 slash 155.

Area reconciliation

The jamabandi gives kanal and marla, the deed gives square yards, the broker gives gaj. Records never agree exactly, so the test is a band rather than equality.

Ask about any term, record or number.

Answers come only from the ontology, the scoring model and the source registry, so nothing here is invented. Arithmetic questions such as converting an area or walking a khasra lineage are computed live rather than looked up.

Topics

Teach it to score the way you score.

Score a file by hand and write your reason against each parameter. The agent solves for the weights that reproduce your judgement, then tells you where they differ from the framework you wrote. That gap is usually the most interesting thing on the page.

2
files you have scored
n/a
leave one out error, points
n/a
verdict band agreement
3
files before it stops guessing

Where your weighting differs from the framework

Positive means you care about it more than the document says you do. Below three labelled files for an asset class the agent keeps the framework weights and says so rather than inventing a preference it has not seen.

ParameterCategory FrameworkYoursDeltaReading

Files it has learned from

PropertyAsset classParameters rated Your overall

How the loop runs

You score, it watches

Open the dashboard, set every slider the way you would on a real file, and press Teach the agent this file. Your reasons go with the scores.

It solves for your weights

Non negative ridge from your parameter scores to the overall you settled on. Non negative because a parameter cannot make a property better by being worse.

It reports its own error

Leave one out, in points out of 100, and how often it lands in your verdict band. A model that cannot say how often it is wrong has no business scoring anything.

It writes in your words

Your own sentences are stored against each parameter and offered back when a new file scores similarly, so a generated card does not read like a template.

Every enquiry, answered from one inbox.

The website form emails contact.tasalee@gmail.com. Paste an enquiry in here and the assistant drafts a reply out of the ontology, the scoring model and the source registry, so nothing in the answer is invented. You edit it, send it from that inbox, and the client's reply comes back to the same thread.

Client enquiries
contact.tasalee@gmail.com The address on the website. Replies go from here.
Ask the agent by mail
supportagent.tasalee@gmail.com For when you are away from this console and want the agent's read on something.

Queue

Nothing loaded

How this is meant to be worked

The inbox is the system

Every enquiry from the website arrives as an email at contact.tasalee@gmail.com with the client's own address as reply to. Answer in the thread and the client sees a normal conversation, not a ticket number.

The draft is a starting point

The assistant answers only from what is mapped, and marks anything it could not find rather than filling the gap. Read every draft before it goes. Your name is on it, not its.

Never quote a score you have not run

A first reply says what we will check and what it costs. It does not estimate a score, a price or a risk from an address alone, because that is exactly the guessing the whole business exists to replace.

The queue lives in this browser

What you log here stays on this machine. It is a working desk, not a shared CRM. Connect the backend and the same screen fetches the inbox and files replies against the property instead.