How HVAX processes your bid documents, integrates with the Anthropic Claude API, structures its outputs, and enforces execution bounds.
HVAX is a risk-extraction pipeline purpose-built for commercial electrical bid packages. It ingests MEP project manuals (specifications), construction drawings, and addenda, then systematically cross-references them to surface scope gaps, specification conflicts, and missing coordination items before the bid deadline.
The system is not a general-purpose chatbot, a takeoff tool, or an estimating calculator. It does one thing: find what the estimator might miss in a 900-page document set, and cite the exact source page.
| Component | Technology | Purpose |
|---|---|---|
| Document Ingestion | Python, PyMuPDF | PDF parsing, text extraction, page-level indexing |
| Topic Router | Python, regex + heuristics | Splits the package into Division-level topic segments |
| Extraction Engine | Anthropic Claude API | Evidence extraction, conflict identification, RFI drafting |
| Verification Pass | Anthropic Claude API | Adversarial second pass that attempts to disprove each finding |
| Structured Output | JSON schema validation | Enforces citation, priority, and action fields on every finding |
| Human Review | Founder review | Every Priority-1 finding is manually reviewed before delivery |
Your bid documents flow through a strict, linear pipeline. At no point does your data leave HVAX's own infrastructure except for structured API calls to Anthropic.
| Hop | What is sent | Encryption |
|---|---|---|
| Browser → HVAX | Raw PDF files | TLS 1.3 in transit, AES-256 at rest |
| HVAX → Anthropic | Page-level text segments filtered by topic — never the full document in a single request | TLS 1.3, Anthropic commercial API (standard 30-day abuse monitoring retention) |
| Anthropic → HVAX | Structured JSON findings with citations | TLS 1.3 |
| HVAX → Browser | Final risk report (HTML dashboard) | TLS 1.3 |
Primary subprocessor: Anthropic (Claude API) for document analysis, alongside Clerk for user authentication. HVAX does not send your documents to OpenAI, Google, or any other LLM provider. We use Anthropic's commercial API tier, which does not use customer inputs for model training. See Anthropic's data usage policy.
HVAX uses the Anthropic Messages API to make structured, topic-scoped requests against your bid package. Here is exactly how:
| Pass | Model | Reason |
|---|---|---|
| Extraction | claude-3-5-sonnet-20241022 | Strong reasoning over long technical documents; 200K token context window handles large specification sections |
| Verification | claude-3-5-sonnet-20241022 | Adversarial checking pass uses the same model to attempt to disprove each finding |
The full bid package is never sent as a single prompt. Instead:
HVAX uses Anthropic's commercial API. Under Anthropic's commercial terms, inputs and outputs from the API are not used to train Anthropic's models. HVAX does not fine-tune, distill, or train any model on your bid data.
Every finding produced by HVAX conforms to a strict JSON schema. This ensures consistency, traceability, and machine-readability across reports.
| Field | Type | Required | Description |
|---|---|---|---|
| finding_id | string | Yes | Unique finding identifier |
| priority | int (1–3) | Yes | 1 = Critical price/scope impact, 2 = Needs estimator review, 3 = Informational |
| category | enum | Yes | One of: specification_conflict, drawing_conflict, scope_gap, addendum_conflict, coordination_gap |
| claim | string | Yes | Plain-language description of the issue |
| evidence[] | array | Yes | At least one source with document, section, page, and verbatim quote |
| addendum_check | string | Yes | Whether a later addendum resolves this finding |
| verification | string | Yes | Result of the adversarial verification pass |
| impact | string | Yes | What changes if this finding is real (e.g., price, scope, schedule) |
| suggested_action | string | Yes | Recommended next step for the estimator |
| rfi_draft | string | No | Pre-drafted RFI question ready for submission |
HVAX operates within strict, documented boundaries. These are not theoretical limits — they are hard constraints enforced in code.
| Bound | Limit | Reason |
|---|---|---|
| Max document size | 2,000 pages per bid package | Quality degrades beyond this; we recommend splitting into trade packages |
| File types accepted | PDF only | PDF is the universal format for construction bid documents |
| Scope | Commercial electrical (CSI Divisions 26, 27, 28 + cross-references to 01) | Prompts, heuristics and validation rules are tuned for electrical scope only |
| No internet access | The pipeline cannot browse the web, access external databases, or call third-party APIs besides Anthropic | Prevents data exfiltration and ensures reproducibility |
| No file system writes | Claude cannot write to disk, execute code, or modify uploaded documents | Read-only access ensures document integrity |
| Timeout | 48-hour maximum processing window per bid package | Aligns with typical bid-day turnaround requirements |
| Data retention | Documents purged 30 days after report delivery (configurable) | See Security page for full retention policy |
HVAX does not load Google Analytics, Meta Pixel, Hotjar, Mixpanel, Amplitude, or any other third-party tracking script on the marketing website or inside the product dashboard.
| Data | Collected? | Purpose |
|---|---|---|
| Client-side analytics (GA, etc.) | No | — |
| Ad-network pixels | No | — |
| Server access logs | Yes | Security monitoring, error debugging |
| Authentication events (Clerk) | Yes | Account creation, sign-in, sign-out timestamps |
| Pipeline execution logs | Yes | Processing time, error rates, finding counts (no document content) |
| Document content for training | No | Your documents are never used to train or fine-tune any model |
HVAX is a research and extraction tool, not an estimator, code consultant, or contractual advisor. These are the things it explicitly does not do:
✗ Not a takeoff tool
HVAX does not count fixtures, measure conduit runs, or produce quantity estimates.
✗ Not a code checker
It does not verify NEC compliance, local code amendments, or AHJ requirements.
✗ Not a legal advisor
Findings are not legal opinions. Contract interpretation must be done by qualified counsel.
✗ Not infallible
AI can miss conflicts and can hallucinate findings. Every P1 finding is human-reviewed, but the tool is not a replacement for your estimator reading the documents.
✗ Not a general AI assistant
You cannot ask it arbitrary questions. It only runs the pre-configured extraction pipeline against your uploaded documents.
✗ Not multi-trade (yet)
Currently tuned for commercial electrical scope only. Mechanical and plumbing trade support is not available.
HVAX surfaces potential issues. It does not guarantee completeness. Your estimator must independently verify every finding against the original documents before incorporating it into a bid price. See our Terms of Service for full liability limitations.
HVAX is designed as a human-in-the-loop system. AI handles evidence extraction at scale; humans handle judgment.
| Stage | Actor | Responsibility |
|---|---|---|
| Document ingestion | Automated | Parse, segment, and index. No human judgment involved. |
| Extraction & verification | Automated | Claude identifies potential findings and an adversarial pass attempts to disprove them. |
| Priority-1 review | Founder | Every P1 finding is manually reviewed by the founder before delivery. Borderline items are separated. |
| Final decision | Your Estimator | The estimator checks each citation against the source document and makes the final call on whether to adjust the bid. |
This is a deliberate design choice. The most dangerous failure mode in AI-assisted estimating is false confidence — an estimator trusting a finding without checking the source. HVAX mitigates this by making the citation the centerpiece of every finding, not the AI's conclusion.