BidVisory
Property owners and managers compare inconsistent contractor bids without a repeatable way to defend the decision.
A source-aware evaluation workflow that extracts competing proposals, verifies arithmetic, researches vendors and market context, and composes an owner-ready report.
My role · Solo founder
- •Problem research and product strategy
- •Product and report design
- •Full-stack and AI workflow engineering
- •Deployment, privacy, cost, and operations
Read transcript
- 01BidVisory starts with the source proposals, including text PDFs, Word documents, scans, and phone photos.
- 02Parallel extraction normalizes scope, prices, dates, terms, and vendor identity.
- 03Research is restricted to eligible source URLs while arithmetic and scoring rules stay in code.
- 04The result is a navigable, owner-ready comparison with red flags and questions to ask.
- 05Cost, latency, structured outputs, source URLs, and evidence links are monitored continuously; field accuracy still requires labeled review data.
- Structured bids
- 25
- Mean vendors
- 2.24
- Expected report cost
- $1.35
- Text-PDF extraction
- $0.009
Retained validator-conforming records
Per report
Current rates and observed mix
One-page measured document
01 / Problem and workflow
From proposals to a defensible decision
BidVisory was designed around the decision artifact, not a chat box. The report must show its work, expose gaps, and survive scrutiny.
Before
Collect PDFs, Word files, scans, and scope documents from vendors.
Read line items and normalize scope manually across inconsistent proposal formats.
Search vendors, licenses, reputation, references, materials, and local market pricing.
Draft a side-by-side comparison, recommendation, red flags, and follow-up questions.
After
Upload the proposal pack without filling a long intake form.
Inspect structured extractions and vendor identities before research is composed.
Receive one consistent report with normalized scope, sourced context, and deterministic math.
Use the report as an owner or board packet instead of rebuilding a spreadsheet.
Four-stage durable workflow
- 01
Extract
Parse each proposal independently into structured scope, price, terms, timeline, warranty, and vendor fields.
- 02
Identify
Resolve vendor identity from the bid and Places data before web research begins.
- 03
Research
Collect market and reputation evidence with an eligible-URL set retained for downstream citations.
- 04
Compose
Generate a normalized report after deterministic arithmetic, scoring, and integrity checks.
Code-owned arithmetic
Totals, averages, adjustments, and reconciliation are computed and tested outside the model.
Fixed scoring rubric
Criteria live in versioned application code so the evaluation posture is explicit and repeatable.
Source URL allowlist
Research claims may cite only URLs gathered in the research stage; unknown URLs are removed.
Retries and fallbacks
Transient failures retry at the step boundary; exhausted optional research degrades explicitly.
02 / Shipped report
The output is the interface
The report keeps the recommendation, comparison, reputation evidence, market context, warnings, and vendor questions in one navigable document.
No-login sample report
Hiring managers can inspect the actual report anatomy without creating an account. Sample vendors and prices are clearly labeled illustrative.
Recommendation hierarchy
The report leads with a recommendation and rationale, then moves into comparable bids, tradeoffs, warnings, and follow-up questions.
Architecture choices
Extraction runs in parallel by document; deterministic normalization and arithmetic happen before report composition.
Vendor identity combines proposal evidence and Places data. Research returns an explicit eligible URL set used by the composition prompt and post-generation validation.
Durable workflow steps retry independently. Completed journals are cleaned because they contain extracted customer document text.
- 01
Documents
Text PDFs, scans/photos, and DOCX proposals.
- 02
Parallel extraction
Independent parsing avoids one bad file killing the entire pack.
- 03
Normalize and verify
Units, totals, dates, and arithmetic are checked in code.
- 04
Identity and research
Places plus allowlisted web sources establish context.
- 05
Compose and validate
The report is assembled, scrubbed, and stored as the durable artifact.
- 06
Privacy cleanup
Completed workflow journals are removed after the durable report is stored.
03 / Evaluation
Measured & Monitored
Recorded usage supports the cost model. Operational telemetry is captured continuously, and the workflow is reviewed and adjusted when models, prompts, pricing, or behavior change.
Measured unit economics
Mean vendors per report
17-report production snapshot
2.24
Text-PDF extraction
Per one-page measured document
$0.009
Three-vendor fixture
Nine AI calls + three Places calls
$1.56
Expected report cost
Modeled from the current vendor mix
$1.35
90th percentile report cost
90% of modeled reports cost this or less
$1.74
Reputation research
Share of measured three-vendor run
59%
Monitored
These signals describe completion, latency, and evidence controls. They do not represent semantic extraction accuracy.
Report latency
90% of 17 retained reports finalized within this time
2m 43s
Structured bid records
Validator-conforming extractions retained; 3 drafts excluded
25
Stored source URLs
HTTP(S) URL present; not claim-level citation coverage
1,321 / 1,321
Rating evidence links
Stored rating entries with an evidence or profile link
57 / 57
Pipeline guardrail tests
Identity, provenance, arithmetic, scoring, and fail-closed behavior
127 passing
Field-level extraction precision and recall, unsupported-claim rate, expert agreement, and human correction rate still require labeled review data before publication.
Decisions and rejected alternatives
Starter allowance moved from 10 to 8 reports and Pro from 40 to 25 after measured cost logs exposed the full-utilization tail.
Vendor-count metering was rejected because the customer does not control how many vendors bid the job.
Reputation caching was rejected by owner decision; freshness and fairness outweighed the largest theoretical saving.
Workflow journals are cleaned after completion because intermediate payloads contain customer document text.
What the sample changed
In the July usage fixture, only 4% of 25 extracted bids printed a vendor website and 8% printed a phone number.
Places matched 5 of 35 vendor observations in that fixture, causing the more expensive profile-discovery fallback more often than assumed.
Applying current rates to the September report mix produces a model of approximately $0.21 plus $0.51 per vendor: $1.35 expected, with 90% of modeled reports costing $1.74 or less.
The next quality claim requires a scored benchmark; this portfolio does not substitute architecture language for that evidence.
Next case study
LineStriper AI
Parking-lot striping contractors lose time moving from aerial takeoff to estimate to signed quote.