Inspiration
Ask a manufacturer what a product or prototype build actually cost, and then ask them to prove it. Engineering BOMs, purchasing records and project hours use different part IDs, units and revisions (MT-250 "1 each" in the BOM, SKU 88-41 "box of 100" in procurement, a blank project field in timesheets). So finance rebuilds cost in spreadsheets, and even then a total that matches but can't be explained doesn't hold up. The problem isn't the arithmetic, it's trust.
We built Bowerbird for Kearney × PERLab Hackathon, Corporate Track 3: BOM Cost Linkage. Our tagline is the goal: Know what a build cost. Prove it.
What it does
Bowerbird links every part a manufacturer designs to every dollar it spends, then opens every number back to its source records.
- Match: spend lines are scored against BOM parts on category, specs and text. 0.85 and up auto-links, 0.50 to 0.85 goes to a human review queue, and below 0.50 is never allocated. Nothing is silently guessed.
- Review: a person confirms or rejects near-misses.
- Allocate: shared parts are split by BOM quantity × build volume under a versioned rule, not an even split.
- Evidence drawer: every number opens to the BOM revision, the PO/receipt/invoice line, the unit conversion, the rule version and each human decision.
- Savings ladder: a raw price gap is kept separate from comparable, negotiated and realized savings.
Results on our synthetic example data:
- 98.3% of $1.01M spend ties to BOM parts; the rest is explained (warranty, service spares, a test fixture).
- One shared motor is split $242,000 / $9,075 between the production run and the pilot build ($235.81 and $353.79 per bike).
- A $113.10 raw bolt price gap across two suppliers, kept apart from realized savings.
We also propose an R&D evidence mode (Proposed, not built): the same engine applied to prototype builds, adding project labor, with each input marked actual, estimated or missing. It produces records for specialist review and makes no eligibility or credit claims.
How we built it
- Working prototype (single web page) with matching, review queue, allocation and evidence drawer, verified on a 12-line synthetic example.
- A build spec with phased acceptance tests, so the numbers have to tie out before moving on.
- Velocity AI shaped the plan: we brainstormed with the research feature, validated the idea with value chain and competitive landscape analysis, built simulated interview personas (procurement buyers, design engineers and other stakeholders), and mapped the regulatory and compliance landscape. That produced 5 strategy canvases, 5 simulated personas used as hypotheses, and 12 interview roles mapped for real customer calls.
- Fixed output label: results are always "allocated direct-material analysis", never "actual cost" or "COGS".
Challenges we ran into
- The surface is crowded. Several platforms already show BOM-linked spend views, so we focused on what's harder to find in public material: scored matches with review, usage-based shared-part allocation, and an evidence trail behind every number.
- Real-world messiness: pack vs. each units, superseded drawing numbers, revision effectivity dates, and purchased not equalling consumed.
- Avoiding false confidence: false matches create false savings, so savings count only from linked, comparable lines.
- Keeping the cost view separate from the incentive view so we never imply tax advice.
- Honest limits: our interviews are simulated, so they're hypotheses, not findings. We have not yet measured match accuracy on real data.
Accomplishments that we're proud of
- Numbers that tie out, with every unmatched dollar explained.
- A matcher that never silently guesses.
- Shared parts split by usage, with every number traceable to source records.
- A bottom-up market sizing with labeled assumptions: about $394M TAM, $236M SAM (range $80M to $522M), and a $3.5M to $9.5M SOM path from year 3 to year 5.
- Labeling everything as Built, Proposed, Synthetic, Vendor-reported or Illustrative, so judges can see what's proven and what isn't.
What we learned
- Finance and engineering don't need a prettier dashboard, they need to defend a number.
- Small savings matter: a 1% materials saving on $31M of materials is about $310K a year (an assumption), and one vendor reports 5% on analyzed spend in six months (vendor-reported, directional only).
- A services-led, file-based pilot is a realistic first step before a subscription.
- R&D credit evidence is gathered at the part, purchase and labor-record level, while eligibility is judged at the business-component level, which is why we keep evidence and conclusions separate.
What's next for BowerBird
- Next 90 days: 12 to 15 customer interviews, then a fixed-fee, file-based pilot on real data (one product family, one closed period).
- Decision gate: continue if 6 of 12 interviews describe the problem and 3 companies agree to a pilot.
- Prove it: a larger messy dataset with precision, recall and false-match rate, plus hands-on demos against incumbent tools.
- Build labor and R&D evidence as customers need them.
- Business model (base-case assumptions, to validate): pilot, then a $30K to $100K a year subscription, then an optional $25K to $40K a year R&D evidence pack.
- Roadmap: ERP/PLM connectors, BOM revision history at scale, and optional human-confirmed AI suggestions for low-confidence matches.
All example figures come from synthetic data. Market and pricing figures are assumptions to be tested with real customers.
Inspiration
What it does
How we built it
Challenges we ran into
Accomplishments that we're proud of
What we learned
What's next for BowerBird
Built With
- claude
- velocityai
