Introducing backfin buy side
Your thesis, and what happened next.
Write down why you hold a company, in your own words. It becomes the assumptions it rests on, and every filing that arrives afterwards is read against them. You hear the moment one is touched. It never tells you what to do.
Margin expansion holds. EBITDA margin reaches about 25% in FY27. Revenue grows high single digit. No generic entry on Tolvaptan or Mirabegron before FY28. Pithampur remediation closes without an import alert.
EBITDA margin reaches about 25% in FY27.
Revenue grows high single digit.
No generic entry on Tolvaptan or Mirabegron before FY28.
Pithampur remediation closes without an import alert.
1,975companies · statements as of any date
143checklist items · nine frameworks
18sector operating-KPI libraries
5 minworker cadence on the live filing stream
Margin expansion holds. EBITDA margin reaches about 25% in FY27.
You wrote it down.
Every filing is read against it.
A worker polls the live filing stream every five minutes, not on a schedule someone has to remember. Most filings pass through dark. When one bears on an assumption you hold, that assumption alone answers, with the quote that drove the call and the filing attached.
LUPIN · the filing stream · classify_evidence_impact
impact ∈ {supports, weakens, invalidates, needs_review} · key signal: margin · a directional call under 55 is demoted to needs review · a quote not copied verbatim from the filing is discarded · thesis health 70 → 66
Promises
What was said. What was filed.
A guidance line is captured verbatim from the call with the date it was filed and held open. When the number that settles it is filed, one to four quarters later, the analyst marks it delivered, delayed or missed against that filing. Nothing is marked kept until a filing says so.
“We would like to reiterate our earlier guidance of high single-digit revenue growth for the organization with EBITDA margins of around 25% during the year.”
Vinita Gupta, CEO, LupinQ1 FY27 earnings call · 7 August 2026 · page 5 of 18filed with NSE 12 Aug 2026, 17:47
Ramesh Swaminathan · page 10
“We are being a little more cautious … the EBITDA margins would be in the range of 24% to 25%.”
Same page as the guidance · bears on assumption 03
“…the overall sales in the U.S. will be impacted by competition in key products like Mirabegron and Tolvaptan this year.”
Checks that run anyway
Scored by the business it is actually in.
The computable items from nine frameworks run on every covered company without being asked, each with the number it came from. A metric the filings do not carry is unknown, never a pass. And a manufacturer’s rules once flagged a wealth manager for an implausible 65% margin, which is simply what that business earns, so financials are never scored by that lens.
LUPIN · manufacturing lens
margin · leverage · cash conversion · dilution
HDFCBANK · bank lens
basis: class-correct metrics from filing XBRL, not the manufacturing lens
In short
It produces evidence. It never produces a recommendation. The judgment stays with the analyst.
By construction
It never tells you what to do.
Buy / Sellthe one output it is built not to have
Filedthe only thing a signal can settle against · never a price, never a consensus
thesis health 66 · starts at 70 · moved only by evidence the analyst has not rejected · a number to argue with, not to act on
Every signal is settled by a filed number. A quote the model did not copy verbatim from the filing is discarded before it is shown.
Written into the forensic framework itself: “a research checklist, not a fraud accusation engine.” Outputs say risk flag or needs review, and a high-severity flag triggers deeper review, not a decision.
Promise settlement against the filed number is the analyst’s today. There is no structured falsifier on an assumption yet, and no accuracy figure is published for the classifier until its evaluation has been run.
The library
Nine frameworks. 143 items.
Each drawn from a named body of published investing and forensic-accounting work. The computable items run on every covered company; the rest are surfaced as questions for the analyst, never scored. Plus 18 sector operating-KPI libraries, every claim carrying the management quote it was read from and its page.
Forensic fraud checklist
Cash conversion against profit, dilution, goodwill against net worth, dividend behaviour under stress.
two critical itemsFinancial statement analysis
Revenue quality, margin drivers, balance-sheet stress. The largest single library.
Unified forensic-alpha
Earnings quality and financial shenanigans; receivables and inventories against sales.
O’glove · SchilitManagement quality
Owner orientation, capital returned against raised, auditor and related-party signals.
Behavioural
Where a thesis is most likely to be fooling its author, as questions the analyst answers.
analyst-read · no filing triggerBusiness quality and moat
Return on capital through cycles; cash backing of reported profit.
Capital and credit cycle
Capex against depreciation, leverage into a downturn, position in the cycle.
Valuation
Expectations embedded in the price against what filings support.
never a target priceCapital allocation
Buybacks, dividends, acquisitions, and whether shareholders were diluted to pay for them.
Where it sits
Not a screen. Not a feed.
The other ways a desk watches a position: a price alert that fires on a move already priced in, or a filing feed you read yourself and hope you do not miss the sentence that matters. Here is where each one stops.
What it does not carry, by design: broker estimates and consensus, price targets, intraday ticks. If it was never in a filing, it cannot settle a signal.
Setup
Hear it where your desk is.
Write the thesis once in the web app. Alerts arrive on Telegram with the filing attached, and the same monitor is reachable over the API for a desk that runs its own tooling. Hosted, in development, access by request.
Next on the bench
Built in the open.
Not shipped yet, listed anyway, so everything above stays exactly what the monitor does today.
Last year’s guidance checked against this year’s filed number automatically: met, missed, or quietly dropped.
A filed number that breaks an assumption breaks it without a model in the loop.
Today a filing on a six percent position and the same filing on a watchlist name are indistinguishable to the alert engine.
Design-partner access
Your thesis, and what happened next.
For funds, PMS and family offices. Send a thesis you already hold, two or three sentences, and we return the assumptions it extracts and what the last four quarters of filings did to each one.
Get accessContact
Siddhi Nagvekar
backfin · design-partner access
Contact
Shreyas Mashelkar
backfin · design-partner access