A 150-page Integrated Annual Report lands in your inbox. Somewhere inside it is a P&L, a subsidiary table, a risk register, a paragraph of management outlook that matters more than the other forty, and a share price sitting in a database that has nothing to do with the PDF at all. Turning that into one coherent, decision-useful report — fundamentals reconciled against filings, technicals pulled fresh, every number cross-checked against every other number — is a half-day-to-full-day job for a competent analyst. We wanted to know what happens when an agentic research engine does the same job, end to end, and whether the result would actually hold up to scrutiny.
We ran the test on Dixon Technologies (NSE: DIXON), India's largest electronics manufacturing services company, using its FY2025-26 Integrated Annual Report as the source document. What follows is not a polished demo — it is the actual working session, mistakes included.
Contents
Section 01 · The Four Layers
The Four Layers
Every Agent Adda deep-research report moves through the same four passes. None of them is "the AI part" in isolation — the value is in how tightly they're wired together.
AI Synthesis + Search
Reads the source filing in full (chunked across a 150+ page PDF), pulls live web search for analyst targets, broker views, and anything published after the filing date, and drafts the narrative — the Investment Read, the Decision Frame, the FY27/28 forward view.
Grounded Fundamentals
Every P&L line, subsidiary PAT figure, and JV detail in the report traces back to a specific page of the Board's Report. Where the source document doesn't cover something — balance sheet, cash flow, quarterly trend — the report says so explicitly instead of interpolating.
Live Technical Engine
Queries the platform's own PostgreSQL market data directly — 514 real EOD sessions — and computes SMA20/50/200, RSI(14), MACD(12,26,9), ATR(14) and volume ratios from scratch, then renders them as charts drawn to that data, not a stock photo of a candlestick chart.
The Reviewer
A pass dedicated to catching what the other three got wrong or left stale — price updates that didn't propagate everywhere, template placeholders, and gaps that got glossed over instead of flagged.
Section 02 · Layer One — Reading So You Don't Have To
Layer One — Reading So You Don't Have To
The source PDF was 13 MB and too large for a single fetch, so the engine pulled it in twenty-page windows — corporate overview, business verticals, ESG, governance and risk, and the Board's Report financials — and carried the extracted figures forward as structured facts rather than a rough paraphrase. That is how the report ended up with Dixon's full FY26 consolidated P&L reconciled to the lakh (₹48,87,280 lakh revenue from operations, ₹2,57,956 lakh EBITDA, ₹1,64,425 lakh PAT), a twelve-entity subsidiary table with individual PAT growth rates, and an eleven-item board-disclosed risk register — not a summary of a summary, but the primary numbers themselves, with the page reference attached.
The test we set for this layer: could we cite a page number for every material figure in the report? For the P&L, subsidiaries, and risk register, yes — Board's Report, page 130 onward. For balance sheet and cash flow, the honest answer was no, we hadn't pulled those lines — so the report says that in plain language instead of quietly filling the gap.
Section 03 · Layer Two — Grounded, Not Guessed
Layer Two — Grounded, Not Guessed
The riskiest habit in AI-assisted research is filling a gap with something plausible instead of admitting the gap. We tested this directly by asking the engine to check whether any exchange filing existed for Dixon more recent than the Annual Report itself.
It queried the platform's live corporate-events, insider-alert, and bulk/block-deal feeds directly — tables that carry 1,030, 1,985, and 2,208 rows respectively across 930 different NSE symbols, with entries dated as recent as 19 October 2026. Dixon had zero rows in all three. Rather than reporting that as "no data available" — which reads like a missing query — the engine first ruled out a symbol-format mismatch, confirmed the feeds were genuinely live and populated for hundreds of other names, and then reported the actual finding: no board meeting is currently scheduled for Dixon in the feed, no insider disclosure has posted, no bulk deal has printed recently, as of that exact moment.
That distinction — a verified negative result versus an unqueried gap — is the entire point of a grounded fundamentals layer. It is also the difference between a report you can defend in a room and one you can't.
Section 04 · Layer Three — The Technical Engine Talks to the Database, Not a Chart Library
Layer Three — The Technical Engine Talks to the Database, Not a Chart Library
Most AI-generated "technical analysis" is a paragraph of hedged language attached to no actual chart. We wanted the opposite: pull the real OHLCV history, compute the indicators from first principles, and draw the chart from those exact numbers.
The engine queried 514 sessions of Dixon's EOD price history straight out of the platform's market.equity_eod table, computed SMA20/50/200, RSI(14), MACD(12,26,9) and ATR(14) in a pandas pass, and rendered the results as inline SVG — a price chart with moving-average overlays and volume, plus RSI and MACD panels — scaled to the real min/max of that specific series. The numbers that came back told a real story: a 52-week high of ₹18,471 in September 2025, a crash to a 52-week low of ₹9,600 by March 2026, and a 50.5% recovery back to ₹14,445 by the report date — with the last two weeks of that rally stalling almost exactly at the 20-day moving average on below-average volume.
That last detail is not something you'd invent. It came out of the data, and it changed the technical read from "breakout" to "recovering base, not yet confirmed" — a materially different, more cautious conclusion than a generic bullish gloss would have produced.
Section 05 · Layer Four — The Reviewer That Catches Its Own Mistakes
Layer Four — The Reviewer That Catches Its Own Mistakes
Midway through the session, the price used throughout the report was updated from a slightly stale ₹14,650 (28 August close) to a fresher ₹14,445 (3 September close), pulled live from the database. Nine separate places in the document referenced the old price, the old trailing P/E, and the old implied upside to the broker target — the hero panel, the metric tiles, the valuation table, and one paragraph in the Broker and Market View section that had been drafted earlier and never touched again.
The reviewer pass caught eight of those nine on the first sweep. The ninth — the broker paragraph, still quoting a "~13% implied upside" and "trailing 54×" that no longer matched the rest of the report — surfaced only when a human reviewer read the live page and flagged the inconsistency back. The engine fixed it, then ran a second sweep across the whole document for the same class of error, and confirmed nothing else was stale.
We are showing this specific miss on purpose. A research engine that only ever demonstrates its successes is not trustworthy; one that catches, names, and fixes its own inconsistency — with a human in the loop as the final backstop — is the actual product.
Section 06 · The Honest Disagreement We Left In the Report
The Honest Disagreement We Left In the Report
The platform runs two different technical classifiers against every stock — a momentum/relative-strength model and a stricter Weinstein-stage model. For Dixon, on the same day, they disagreed: the momentum model read STRONG_BULLISH / BUY; the stage model read STAGE_1 / NEUTRAL.
Most report-generation pipelines would pick one and move on. We put both in the report, side by side, with the reconciliation spelled out: the momentum model is right that price is up 50% off its low and sitting above both key moving averages; the stage model is also right that Weinstein's Stage 2 requires a volume-confirmed breakout above resistance, and the last two weeks show the opposite — a stall on light volume. Read together, the two disagreeing signals say something a single confident number never could: strong recovery, not yet confirmed. That is a more useful sentence than either model alone.
Section 07 · What This Actually Took
What This Actually Took
Building this report end to end — reading the annual report, extracting and reconciling the financials, querying live market data, computing technical indicators from scratch, drafting the narrative, and running the review pass — took a working session measured in tens of minutes, not the half-day-to-full-day a manual equivalent typically requires: reading the filing cover to cover, pulling price history into a spreadsheet, building the charts by hand, and cross-checking every figure against every other mention of it across a 20-section document.
That gap is consistent with what we see across report types on the platform more broadly — internal benchmarking puts the time saved at roughly 50-60% versus a fully manual research workflow, once you count not just the writing but the reconciliation work that normally eats the second half of an analyst's day. The Dixon session is one full worked example of where that time actually goes: not into typing sentences, but into the search-grounding, the database queries, and the review sweep that a rushed manual process is the first thing to skip under deadline pressure.
Section 08 · What We're Not Claiming
What We're Not Claiming
This is not a claim that the process needs no human in the loop — the broker-paragraph miss above is proof it still does, today, and we'd rather show that than hide it. It is not investment advice, and the finished report says so explicitly, more than once. It is not a claim that live data stays fresh forever — every technical and filings-feed reference in the report carries the date it was pulled, because a number without a timestamp is a number you can't trust six weeks from now.
What we are claiming is narrower and, we think, more useful: that AI synthesis, live search, a grounded fundamentals discipline, a technical engine that reads real data instead of describing a stock chart in prose, and a reviewer pass that treats its own output with suspicion, wired together, produce a report that survives being read closely — including by the person who built it.
The full report is live on Agent Adda's Market Intelligence page: Dixon Technologies (DIXON): FY26 Annual Report Deep Research.