Earnings Call Intelligence Engine: From Public Filings to Strategic Briefs
A self-initiated, three-stage LLM pipeline that synthesises competitor earnings calls into auditable strategic briefs. The system is generic, its strategic value lives in a customisable "lens" (a set of dimensions a human strategy team cares about), and a single instruction file controls what counts as signal. Stage one extracts the lens-relevant passages from a verbatim transcript. Stage two writes a strategic bullet for each, anchored to a verbatim CEO/CFO quote. Stage three composes the per-company briefs into one quarterly digest with an executive summary on top. Built independently, no client engagement, no proprietary data, as a working demonstration of how this kind of intelligence layer sits underneath any strategy team. Edition 01 was calibrated for a travel-tech case study (Wego, with Booking Holdings as the Q1 2026 source company), built ahead of an interview conversation; the same pipeline runs against any industry with public earnings calls.
Pipeline Architecture
Public earnings transcript in · a human-defined strategic lens encodes what counts as signal · stage-1 extract pulls lens-relevant passages · stage-2 synthesises strategic bullets anchored to verbatim quotes · stage-3 composes per-company briefs into one digest with an executive summary on top.
What’s actually happening at each stage
Each stage is explained twice, first for the finance reader, then for the engineer.
1. The Skill: Earnings Calls Into Auditable Strategy
Finance lens
Every quarter, public companies hold earnings calls that contain real strategic signal, pricing posture, capital allocation, competitive moves, distribution shifts, regional disclosures. Reading four to six hours of transcript across a five-company comp set is not how strategy teams spend their time. The skill this engine demonstrates is turning that public information into a one-page-per-company strategic brief, in a way a senior reader can verify line by line. The output is calibrated to the strategic dimensions that matter to a specific business; the input is anything publicly said by an executive on the record.
Engineering lens
The architecture is pipeline-shaped on purpose. Inputs (transcripts) are commoditised; outputs (briefs) are scarce; the value lives in the prompt-encoded lens that maps from one to the other. Same machinery, configurable strategic surface, point it at any industry, encode that industry's strategic dimensions, and the same pipeline produces a tailored digest.
2. The Lens: Where the Strategic Work Actually Lives
Finance lens
A generic earnings-call summariser is useless. The work that makes a digest valuable is deciding what counts as signal in the first place. For an OTA the lens is regional exposure, AI distribution, take-rate dynamics; for a bank it would be net interest margin, deposit beta, credit migration; for a SaaS firm it would be net revenue retention, gross margin trend, multi-product expansion. The lens is the only place a human does strategic work in this system; everything downstream is the model executing against it.
Engineering lens
The lens is encoded as a single instruction file with N named dimensions (typically 6–12) and a one-line definition per dimension. Change the file and the entire pipeline produces a different digest, same machinery, different strategic surface. Deliberate: the system's value lives in the prompt design and the strategic taste behind the dimensions, not in any model trick.
3. Three-Stage LLM Pipeline: Extract, Synthesise, Compose
Finance lens
Stage one, extract: the model reads the verbatim transcript with the lens in hand, and returns the passages that touch any of the named dimensions. Stage two, synthesise: each extracted passage is turned into a strategic bullet with two parts, a verbatim quote from the executive, and a one-line implication for the reader's business. Stage three, compose: the per-company bullets are stitched into a single digest with a five-line executive summary on top. The reader gets a structured brief; the source material is auditable from the page.
Engineering lens
Implemented in Python with OpenRouter as the model gateway (Claude Sonnet 4.6). Each stage runs as a separate LLM call with a stage-specific prompt, so iteration on synthesis is decoupled from iteration on extraction. Output is structured JSONL between stages and clean markdown at the end. Cost: ~$0.50 per full edition. Wall time: ~5 minutes.
4. Source Discipline: Why It Doesn't Make Things Up
Finance lens
The biggest failure mode of any AI tool that writes about a business is that it sounds confident while saying things the source never said. The synthesis stage is constrained at the prompt level: every bullet must anchor to a verbatim quote, every quote must come from the transcript, and any bullet that can't produce a real quote does not get written. The structure of the output enforces this, each entry on the page is a quote block plus an implication line. If the quote is missing, you see it on the page immediately.
Engineering lens
No paraphrase mode, no second-hand summary mode, no model-speculation mode. The composer pass flags any bullet whose quote doesn't literally appear in the source transcript and re-runs synthesis until it does. Auditability is not a label on the methodology page, it's a property of the generated artifact.
5. Two Surfaces: A PDF for Forwarding, a Reader for Drilling
Finance lens
A strategy team needs two things: something they can email and forward to executives (PDF), and something they can browse when they want the underlying source (web reader). The PDF is the canonical artifact, four pages, executive summary on top, per-company brief below, methodology and coverage at the back. The reader sits on the same synthesised output, navigable by company, edition, and lens, with the verbatim quote one click away.
Engineering lens
PDF rendered from the same markdown source via Chrome headless with editorial CSS (Lora + Inter, navy accent, quote-card layout). Reader is a lightweight Flask app reading the synthesised markdown, no database, no JS framework, under 500 lines of Python. The markdown source is the single source of truth; PDF and reader are two views over it.
6. Edition 01: Travel-Tech Case Study
Finance lens
To pressure-test the engine end to end, Edition 01 was calibrated for a travel-tech reader and run against Booking Holdings' Q1 2026 earnings call (April 28, 2026). The lens for this edition encodes nine travel-specific dimensions (regional exposure, India outbound, SEA distribution, AI/agent distribution, capital posture, marketing economics, partnership network, funnel conversion, forward guidance). The output: a 5-takeaway executive summary plus 8 quote-anchored strategic bullets, every implication traceable to a Glenn Fogel or Ewout Steenbergen quote. Built ahead of an interview conversation with one travel-tech strategy team; the engine itself is independent of any single company.
Engineering lens
Edition 01 is the proof of concept, generic engine plus a domain-tuned lens. Edition 02 would broaden the panel to the full five-company travel-tech set; alternative editions could re-target the engine at any industry with public earnings calls.
Methodology notes
Lens-as-prompt design: a single instruction file encodes the strategic dimensions a human team cares about. Every downstream stage executes against this list, change the file, and the entire pipeline produces a different digest from the same source material
Three-stage LLM pipeline (extract → synthesise → compose) with stage-specific prompts so iteration on synthesis is decoupled from iteration on extraction; structured JSONL between stages, clean markdown at the end
Source discipline enforced at the prompt and structural level: every bullet anchors to a verbatim quote, every quote must literally appear in the transcript, every claim is auditable from the page, no paraphrase mode, no second-hand summary mode
Two surfaces over one markdown source: a 4-page PDF rendered via Chrome headless with editorial CSS for forwarding, and a Flask reader for drilling into the underlying quotes, single source of truth, two views over it
Cross-domain by construction: the same architecture runs against any industry with public earnings calls or filings, banking (NIM, deposit beta, credit migration), SaaS (NRR, gross margin, expansion), energy (production guidance, capex, hedging), or country-level sovereign credit (the same lens-filter-quote pipeline applied to IMF Article IV reports)
Adjacent applications fall out for free: M&A target screening (point at potential acquisition targets' filings with an integration-thesis lens), partnership ICP profiling (extract partnership signals from 10-Ks), competitive intelligence for any industry where strategy teams currently rely on a junior analyst reading transcripts
What this isn’t (yet)
The honest limits. A page called “honest” with no limitations would be a credibility own-goal.
Self-initiated · no client engagement
This is an independent prototype. It was not commissioned by, paid for by, affiliated with, or endorsed by any of the companies referenced (Wego, Booking Holdings, Expedia, Trip.com, MakeMyTrip, Tripadvisor). All input data is verbatim from public earnings calls; the strategic implications are the author's own analysis.
Edition 01 is one case study
Edition 01 covers Booking Holdings only, chosen because it had just released Q1 2026 results at the time of build. The full five-company travel-tech panel is Edition 02 and depends on the remaining Q1 calls landing.
Strategic taste is in the lens, not the model
The system is only as good as the dimensions encoded in its lens file. A generic or poorly-thought-out lens produces a generic digest. The leverage is the human strategic judgement that goes into deciding what counts as signal, the model just executes against that judgement.