In 2023, the market was flooded with "chat-with-your-PDF" applications. The pitch was intoxicating: just drag a 500-page merger agreement into a chat window and ask, "Are there any risks?"
But for institutional due diligence—M&A lawyers, private equity associates, and investment bankers—this approach quickly proved unusable. Why? Because a general-purpose LLM is fundamentally misaligned with the rigor of legal and financial review.
"A general-purpose AI is designed to be helpful, creative, and conversational. A due diligence analyst needs to be skeptical, exhaustive, and exact."
When you ask a generic AI wrapper to find risks, you encounter three immediate liability problems:
- Boilerplate blindness: The AI will flag standard "stockholder approval" mechanics as massive risks because they sound legally complex, completely failing to separate materiality from standard transaction mechanics.
- Hallucination: When pressed to find a specific clause, the AI will often invent one that sounds perfectly plausible.
- Instability: Asking the exact same prompt twice yields different severities.
We built Anweshna to solve exactly these problems. We do not use the LLM to make the final decision. Instead, we use the LLM strictly as a natural language processing module, wrapped inside a massive, deterministic 7-layer software pipeline. Here is how it works.
1. The Industry-Specific Ruleset Engine
A generic AI doesn't know if it's reading a commercial real estate lease or a pharmaceutical patent license. Anweshna's pipeline starts by injecting a rigid, industry-specific taxonomy into the extraction phase. The AI isn't asked to "find risks"—it is commanded to extract entities matching our precise, proprietary 15-category M&A ontology.
2. Dual Scoring Engines
A 500-page data room dump breaks most AI context windows. Instead of relying on lossy RAG (Retrieval-Augmented Generation) which often misses needle-in-the-haystack clauses, Anweshna uses a Map-Reduce Deep Scan. We chunk the documents, analyze each sector in complete isolation, and then algorithmically reconcile the findings.
3. The Missing-Information Detector
Perhaps the hardest thing for an LLM to do is realize what isn't there. Anweshna maintains a deterministic baseline of required disclosures per industry. If the target company fails to provide an IP schedule or environmental compliance report, our pipeline triggers a "Diligence Gap" flag automatically.
4. Deterministic Evidence Verification
This is where we fundamentally diverge from a ChatGPT wrapper. Every single finding generated by the AI in Anweshna must be verified. Our proprietary evidence_verifier engine strips the AI's response and searches the original, raw extracted text for the exact quoted span or figure. If the AI hallucinated the number or the quote, the engine flags it as [UNVERIFIED]. The AI cannot bypass this gate.
5. The Second-Stage Materiality Verifier
We recently introduced the anweshna_verifier, a second-stage deterministic engine that re-classifies every AI finding. It separates Evidence Type from Severity. If the AI flags a "no-shop provision," the verifier demotes it to a [TRANSACTION MECHANIC]. It ensures that only actual, materialized issues (like a live regulatory action) are escalated to human review, radically reducing false positives.
6. Negation-Aware Risk Detection
Contracts constantly list risks that they have explicitly mitigated. Our pipeline uses a two-tier negation detection system to ensure that a sentence like "The Company is NOT subject to any material litigation" does not trigger a false-positive litigation flag, a common failure mode for basic vector searches.
7. Critical-Phrase Floors & Blocking Gates
Finally, we never let the AI have the final say on deal-blocking severity. We enforce strict "critical-phrase floors." If the document contains the phrase "going concern" or "substantial doubt," our deterministic code immediately overrides the AI's subjective score, forces the category into a Blocked state, and halts the deal until a human intervenes.
The Result: Institutional Trust
By moving the burden of severity assessment, hallucination checking, and category gating out of the LLM and into deterministic code, Anweshna achieves precision rates that generic wrappers cannot match. Our latest benchmark achieved an 84.6% precision rate on deal-blocking flags, with a 0% false-negative rate on known deal-killers.
Anweshna isn't a ChatGPT wrapper. It is a deterministic diligence pipeline that happens to use an LLM for text extraction.