Boom Leverage

The period covers a horizontal expansion of the Thai index and a change in the pricing structure. Two new searchable axes opened (Governance, and Business & competition), and AI transcripts of Opportunity Day briefings were added as a distinct, searchable evidence class. Pricing moved from flat per-action costs to an axis-based formula, and the free daily quota rose from 10 to 100 credits.[13] ALPHA history was extended back to FY2016. Vietnam remains in an early-access corpus-building phase, unmonetised, serving zero companies.

The disclosure regime obliges every listed entity to file extensive qualitative documentation each year. For the Stock Exchange of Thailand, the Form 56-1 One Report consolidates hundreds of pages covering operations, risk factors, corporate governance and management's discussion and analysis. At market scale the qualitative half of these filings remains largely unread by institutional participants. The quantitative elements — net profit, balance-sheet ratios, standard financial metrics — are digitised and available in milliseconds through incumbent terminals; the prose is buried in unstructured, highly variable PDFs. Retail investors rely on delayed third-party summaries, while institutional desks fall back on static keyword search. The bottleneck produces a persistent latency: operational distress signals such as shifting customer concentration, localised supply-chain failure or abnormal receivables ageing are formally disclosed by management quarters before they consolidate into a media narrative or a price correction.

"Unread" rather than "unavailable" is the commercially load-bearing distinction. The alpha in this domain does not come from acquiring proprietary data; it comes from reading existing public data comprehensively and systematically. Literal string search fails consistently against corporate prose because management teams describe distress in euphemism and highly variable vocabulary. An analyst searching the exact phrase "customer concentration" will miss a filing that states "the majority of the Company's revenue recognized… comes from government and state enterprises customers", because the character string is absent. A retrieval system without semantic comprehension does not find the underlying risk.

Boom Leverage occupies this void by converting the Thai market's Form 56-1 filings into a semantic vector space.[1] The institutional workflow shifts from guessing keywords to applying a standardised risk taxonomy across the entire cross-section of the market at once. The value is anchored in corporate memory rather than forecasting: the platform holds the historical baseline of disclosures up to ten years deep on the ALPHA tier, which is what surfaces the quarter a management team changed its story about a specific operational risk.[7] The gap the product fills is the human inability to track semantic change across thousands of unstructured documents over multi-year periods.

The Form 56-1's limitation is that it is a heavily drafted, legally reviewed document designed to limit liability. Management's verbal guidance at earnings briefings carries hesitation, unscripted context and shifts in tone that never reach the printed filing. To reach that layer, the platform extended its scope to the Stock Exchange of Thailand's Opportunity Day briefings: AI transcripts are searchable alongside the filings. When last counted, on 21 September 2026, the rolling four-year window held 5,755 sessions; the window widened to six years on 30 September 2026 and has not been recounted for this edition.[16][7] Separately, an archive of 254,149 verified findings covers FY2016–FY2020. It is reported apart from the Core window and never aggregated with it, so that figures with different denominators stay apart.[4]

The extraction architecture is built on a spatial bounding-box paradigm: a valid result consists of four inseparable components — the literal passage, the specific filing, the exact page number, and the verifiable coordinates of the verbatim sentence on that page. Rather than operating purely in string space, the engine ingests the Form 56-1 PDFs and records their spatial metadata. On a query, a multilingual embedding model retrieves the relevant passage. Where a document is a scanned image with no embedded text layer, optical character recognition and vision models are used strictly to recover word coordinates. The OCR layer is prohibited from altering semantic content; it is an alignment tool and nothing else.

The defining control is the citation gate. Before any PDF result renders, the system runs a deterministic match between the extracted passage and the source text at the recorded coordinates. If the match fails for any reason — formatting corruption, OCR failure or model hallucination — the output is suppressed. As of 30 September 2026 the gate recorded a 98.95% pass rate across the five published axes, 489,178 verified findings out of 494,377, and the withheld count of 5,199 is published beside it, which makes the failure rate auditable rather than absorbed into a general accuracy metric.[2] The served index holds 598,776 rows whose evidence is a verbatim PDF quote, including the beta forensic axis, which is tracked separately at 99.11%.[14][15]

For institutional users who do not read the source language, the architecture carries a non-destructive verification path. Every cited line is presented bilingually: the company's original Thai sentence verbatim, holding its status as the authoritative text, and the English rendering beside it, labelled as a translation. Because the filing link and the page coordinate arrive on the same row, an English-speaking portfolio manager can escalate any flagged sentence to local counsel or a bilingual analyst and have it settled in minutes. The architecture refuses to compel trust in an unverified translated string.

The Opportunity Day transcripts introduced a second evidence class that deliberately sits outside the bounding box. The source is a recording, so the evidence for a spoken passage is a link to the minute of the webcast, not coordinates on a printed page. The two pipelines are kept separate. A product that blended a transcription fill rate into its PDF verification rate would inflate the metric its institutional pitch rests on, so transcript rows are excluded from the citation gate that governs the filings.

The platform operates on a philosophy taken from bank-grade model risk management: a false positive is more destructive to institutional trust than a missing signal. A system that routinely paints benign disclosure as severe operational risk trains portfolio managers to ignore its output. Boom Leverage therefore deprioritises alert volume in favour of signal density. Where an extracted passage cannot meet the bounding box's verification standard, the system fails closed: it withholds the finding rather than interpolating, estimating or averaging around it. Pinning the 5,199 withheld items to the interface forces the system's operational limits into view instead of flattering its parsing accuracy.[2]

The discipline is enforced hardest on the auditor-opinion axis. Because accounting rules oblige every listed entity to disclose standard liquidity and credit risk, blanket extraction produces unactionable noise. The mechanism evaluates the specific wording of the auditor's text rather than triggering on the presence of a mandatory structural header, isolating modified opinions from routine Emphasis of Matter disclosures. The consequence of that threshold: out of 910 companies scanned across FY2021–FY2025, 68 were flagged.[9] The system also quarantines regulatory boilerplate so a structural artefact cannot present as idiosyncratic distress: a deterministic rule quarantined 96 rows tied to the TSA 570 going-concern template, clearing 60 otherwise clean companies from the red-flag register.[11]

The move into AI-transcribed briefings required the same philosophy applied to spoken language. Speech recognition on Thai financial vocabulary gets units wrong with full confidence — "9,200 million baht" can come back as "9,200 baht". To keep numerical errors out of the terminal, a figure extracted from speech must also appear on the presentation deck published for that session, and it is separately magnitude-checked against the platform's financials corpus. A spoken figure that fails either test is not served.[16]

The system also refuses to assign severity scores to spoken transcripts. The filing taxonomy grades written risk disclosures, but Opp Day rows are labelled as machine transcripts and their severity field is null. A machine-assigned judgement of an executive's tone, hesitation or unscripted phrasing is an unexplainable variable, and a score placed over an unscripted conversation would be fabricated authority. The product rejects it outright.

Data integrity rests on the immutability of the source regulatory text; isolation prevents cross-tenant contamination. The primary leak class in semantic vector infrastructure is query visibility — where one institutional tenant's search terminology informs the autocomplete, caching layer or model weights that a competing tenant meets. The platform enforces isolation through a rigid separation between the static, market-public document index and the dynamic query layer. User search histories, applied taxonomies and proprietary queries are excluded from any centralised continuous-learning dataset. The mechanism exists so that a risk hypothesis developed by one desk is not broadcast to the market through a fine-tuned model update.

The second leak class is generative hallucination, typically reached through prompt injection or weak model boundaries. It is mitigated by removing generative synthesis from the core search loop. The architecture is restricted to vector retrieval, passage matching and coordinate mapping, and the interface holds no generative mechanism authorised to summarise, infer or forecast financial outcomes. Adversarial input is processed as a literal semantic vector, matched against conceptually proximate paragraphs, and put through the bounding-box citation gate. Where no verbatim match exists on a verifiable PDF page, the query fails safely, returns nothing, and does not construct a narrative.

The third class is staleness, where a figure validated in one period is presented as current in a later one. Figures frozen to a measurement date are labelled as such rather than refreshed to look current. The rule governs this document too: edition 02's 98.94% over three axes was not overwritten by this edition's 98.95% over five, and the Opp Day session count is shown with its own date and a note that the window has since changed. Both gate figures remain readable side by side, bound to their populations and dates.[2][16]

What remains unevidenced is set out in the open questions of this report. The evidence pack contains no external attestation: no SOC 2 Type II or ISO 27001 certification is evidenced, and cloud-level tenant isolation for the ALPHA desk licence — dedicated VPC allocation or separate tenancy encryption — is undocumented. The architectural controls in this section are functional constraints, not a substitute for third-party security validation. An institutional vendor review should proceed on the assumption that this gap is open.

The model is an action-metered freemium subscription ladder designed to remove evaluation friction before converting institutional users to recurring revenue. The entry point is a Google sign-in granting 100 free credits a day with no payment details, limited to the latest filings, about five quarters.[13] DELTA at ฿799 a month extends the horizon to four years of Form 56-1 history and an eight-round Opp Day window, with 150 credits a day.[5] GAMMA at ฿1,399 buys 350 credits, the full six-year Core window and 16 Opp Day rounds.[6] Prices are in baht rather than dollars, which keeps foreign-exchange volatility out of a regional asset manager's budget and keeps the storefront and the checkout quoting the same number.

Usage metering moved from flat per-action prices to an axis-based formula, and the formula states where the product thinks its value sits. Searching the first one to five axes together costs one credit, with one more for each axis beyond five and one more per 25 companies beyond the first 25.[10] The dearest actions are not broad searches but targeted verification: opening a filing page with the quote boxed costs two credits, and revealing one company's auditor findings costs five. Following the citation out to the regulator's own site costs nothing — when the reader checks us against the source, no toll is drawn.[10]

At the top of the ladder, the ALPHA desk licence starts at ฿9,900 a month for 1,500 credits a day, multiple seats, API access and bulk CSV/Excel export, with history back to FY2016 — ten years of corporate memory.[7] The ladder is anchored against a bespoke service, On-Demand Research, from $1,000 per report.[8] Presenting a $1,000 report price beside a ฿799 monthly plan is a pricing mechanism: an analyst reads the subscription as small next to the labour of doing the same diligence by hand.

What this document cannot quantify is the other half of unit economics. The marginal compute cost of a retrieval and coordinate-mapping sequence is absent from this evaluation, so no gross margin per action is claimed. Annual recurring revenue, retention, active customer counts and account concentration are also absent from the evidence pack. That omission leaves the revenue narrative unauditable by a third-party reviewer.

The platform is founded and operated by Varanchai Yingkhamnueng. His background is in quantitative model risk management and banking analytics rather than software engineering. Trained in the hard sciences — an M.Sc. in Physical Chemistry with highest distinction — he moved through quantitative proprietary trading into institutional banking. At Kasikornbank he worked in credit-risk quantitative analytics and built the institution's first natural-language early-warning system, applying deep learning to news and unstructured prose to catch credit-risk signals before they became non-performing loans. At Krung Thai Bank, as a Senior Data Scientist in business risk and macro research, he owned and validated the production monthly expected-credit-loss models the bank used under TFRS 9.

That pedigree dictates the platform's risk posture. A practitioner who has had to defend a TFRS 9 ECL model in front of internal bank auditors acquires a durable distrust of unexplainable variables. The structural refusal of generative synthesis, the false-positive discipline governing alert volume, and the wording-severity taxonomy on the auditor axis are that framework translated into a commercial product. The Thai SEC IC Complex Type-1 investment consultant licence anchors the same adherence on the capital-markets side.

The same fact is the platform's largest unmitigated operational risk, and it is stated here rather than left as an implication. The entity runs on an extreme single-key-person dependency. The technical architecture, the calibration of the risk taxonomy, the bilingual prompt engineering and the go-to-market execution rely on one person. The evidence pack documents no engineering team, no co-founder and no institutional capital. Were the founder incapacitated or to pivot away, the platform's ability to maintain the vector index, adapt its parsers to regulatory format changes, or execute the multi-market thesis would stop. §10 carries it as the first risk.

The expansion thesis is a franchise mechanism: replicate the semantic indexing and bounding-box extraction engine across adjacent Asian markets with a similar centralised, mandatory disclosure framework. Vietnam is the designated second market, targeting the Annual Reports (Báo cáo thường niên) of entities listed on HOSE, HNX and UPCoM. The rationale is lateral scaling: reuse the multilingual embedding models and coordinate-mapping software to parse a new language without redesigning the core search and verification loop.

The Vietnamese integration is in a corpus-acquisition phase. The search infrastructure is live alongside the Thai pipeline and the localised ladder is published, anchored at ₫639,000 a month for DELTA, derived from the ฿799 Thai price at a fixed 1:800 rate.[12] Exactly zero Vietnamese companies are currently indexed and searchable.[3] The gating condition is deliberate: no market opens to paid queries until its corpus clears the same verification threshold enforced on the Thai Form 56-1 pipeline.[2] The early-access waitlist lets clients request priority indexing for named companies, which implies demand-driven processing rather than a market-wide ingestion before launch.

Three conditions have to hold for the thesis to become a business. The formatting of Vietnamese filings must carry a predictable text layer, or remain tractable to the vision models, so the citation gate functions. The embedding model must translate Vietnamese business terminology into operational distress signals as accurately as it does in Thai. And local institutional demand must validate the ₫639,000 price without bespoke product changes that would fork the engine. Until a verified Vietnamese corpus reaches the public interface, the expansion thesis is an unproven capability.

Each entry isolates a structural dependency, a regulatory exposure or a technical bottleneck, and states what would show the concern answered. The open questions above — undocumented recurring revenue, unmeasured compute cost per query, the absence of SOC 2 Type II or ISO 27001 certification, and no published accuracy rate for Opp Day transcripts — inform this register rather than sitting apart from it.

1. 01Key-person concentration

   Mechanism

   The technical stack, the banking risk taxonomy and the go-to-market execution rely on a single founder. Incapacitation or a strategic pivot halts pipeline updates, regulatory adaptation and regional expansion immediately.

   What would answer it

   The firm hires a dedicated engineering lead, or secures institutional capital that enforces operational redundancy and documented knowledge transfer.

2. 02Regulatory format dependency

   Mechanism

   Bounding-box extraction depends on the structural predictability of the Form 56-1. If the regulator materially alters the required taxonomy, the parser may stop classifying items correctly into the established lenses.

   What would answer it

   The regulator moves to raw-text APIs instead of rendered PDFs, which removes the spatial-parsing dependency altogether.

3. 03Translation liability on cited lines

   Mechanism

   The Thai text is preserved as the authoritative artefact, but an institutional client acting on a mistranslated English rendering still takes the loss. The failure survives every control that checks whether the passage exists, because the passage exists and the translation is wrong.

   What would answer it

   A sampled bilingual review of cited translations with a published error rate, repeated on a schedule.

4. 04Vision-model exhaustion on scanned filings

   Mechanism

   Recovering word coordinates from image-only PDFs with vision models carries a compute cost and a failure rate. Every failure is a finding withheld, so the mechanism that protects the reader from false positives also quietly removes real signal.

   What would answer it

   The citation pass rate across image-heavy filings holds above 98% without compute cost rising disproportionately — reported by filing type rather than as one blended figure.[2]

5. 05Commoditisation by an incumbent

   Mechanism

   Established terminal providers hold large distribution moats. If one deploys equivalent semantic vector search natively, the technological edge commoditises immediately, leaving only the risk discipline as a moat.

   What would answer it

   An incumbent ships the feature and its alerts prove too noisy for institutional risk desks, showing the false-positive discipline is a product decision rather than a technical one.

6. 06Transcription error on Opp Day

   Mechanism

   Opportunity Day transcripts are produced by speech recognition and sit outside the bounding-box gate. The slide-deck check governs figures only; a wrong word, name or unit in the surrounding text can still reach an analyst, labelled as a machine transcript but not verified.

   What would answer it

   A published, human-audited word error rate on a statistically sound sample of transcripts, with numerical entities reported separately.

Every quantitative claim in the prose above, bound to the population it counts, the date it was measured, its source and the way a third party re-checks it. A figure absent from this table is unsupported by the evidence pack and has been kept out of the analysis.