Manifest
AI-native procurement agent for hardware direct materials — BOM ingestion, sourcing timelines, RFQ normalization, tariff & lifecycle risk. Co-founder evaluation with Vraj (ex-Infineon / Lucid / Zoox).
Verdict
The idea clears the bar. Real pain × real timing × a real (if narrowing) gap × rare founder-market fit is a strong combination. If someone is going to build the agentic direct-materials procurement tool for hardware startups, the conditions are as good now as they'll be.
But it's a race, not an open field. Winning requires: speed into the 12–24mo window, the agentic + startup-native wedge (not "AI procurement" in general), email-native / no-portal discipline, and converting Vraj's warm network into paid design partners fast. The real question isn't "is the company good" — it is — but "am I the right CTO and is this the bet I want?"
Bull & Bear Case from docs/06-thesis.md
Pain is real, acute, expensive
Direct procurement runs on email and spreadsheets. Missed builds = board-visible. This isn't a vitamin.
Timing is unusually good
Hardware renaissance minting target customers; 2025–26 tariffs = every BOM is a re-sourcing problem; agentic AI just got good enough.
Real gap in the market
LightSource ($33M) and Cofactr ($28.8M) are AI-assisted, not agentic. Lumari (YC X25) is equally early. ~12–24mo window.
Rare founder-market fit
Vraj was the customer (Lucid, Zoox). Warm network of VPs/CPOs/investors. One drone partner verbally committed.
ROI math closes
1% of $30M = $300K; a $50–100K tool is still a 3–6× ROI. ACV is realistic for the segment. (At $300M spend, 1% = $3M — verify Vraj's reference figure.)
Buildable fast
Three wedge prototypes built in one day, on a real BOM, with real tariff data. That's the proof of concept.
Crowded and crowding fast
LightSource, Cofactr, Lumari, Tacto ($50M), Luminovo, Procure Ai — all funded. Automation alone isn't the differentiator.
Thin moats early
The only durable moat is workflow system-of-record, which accrues slowly. Valuable data (prices, lead times) is chicken-and-egg.
Venture-scale only with expansion
Beachhead alone = strong Series A/B, not unicorn. Thesis depends on the land-and-expand into ERP territory.
Cautionary comps are real
Scoutbee raised $76M → Coupa acqui-exit. Flexport marked down ~80%. Procurement software is hard to sell and defend.
Competitive Landscape from docs/01-competitive-landscape.md
Direct / Hardware (the real arena)
| Company | What it does | Agentic? | Funding | Gap |
|---|---|---|---|---|
| LightSource | "AI-Native Procurement for Direct Materials" — BOM, RFQ, award. Customers: Infinite Machine, BRP, Shure | AI-assisted, NOT agentic | $33M (Lightspeed + BCV) | Augments humans; not startup-native pricing |
| Lumari (YC X25) ⚠ | "AI procurement automation purpose-built for direct materials" — always-on agents for RFQs, PO expediting, supplier follow-up | Genuinely agentic | YC Spring 2025 (early) | Very early; concept-clone; the one Vraj didn't mention |
| Cofactr | Electronics procurement + ITAR warehousing/kitting; 100M+ parts; acquired Factor.io + Cogbase | Hybrid: AI + asset logistics | ~$28.8M (Bain + YC) | Asset-heavy; ITAR/defense focus; 3PL-with-software |
| Tacto | AI "OS" for German Mittelstand: supplier mgmt, ESG, sourcing; 250+ named agents claimed | AI-assisted → agent-positioned | $50M (Sequoia, Index) | Compliance/EU framing; not US hardware-startup BOM-first |
| Luminovo | EU electronics procurement & quoting; ElectronicsGPT assistive chat | AI-assisted workflow | ~$16M seed | EU/EMS-centric; quoting-side bias |
| CADDi | AI data platform for machined/custom parts (drawings → procurement) | AI data + analytics | $232M @ $470M val | Mechanical/custom parts; not electronics BOM RFQ |
The "$1.5M Amazon Lab126" competitor (Vraj's unnamed)
Best-supported hypothesis: LightSource. Tagline literally matches; demonstrably non-agentic by own marketing; funding profile ($33M, >$1B processed, large enterprise customers) consistent with a ~$1.5M contract. The Lab126 relationship and $1.5M figure are Vraj's insider data — not independently verifiable. If Vraj can confirm, this validates the pain and the non-agentic gap simultaneously.
Indirect / S2P (not our competition)
| Company | Funding | Why irrelevant |
|---|---|---|
| Zip | $190M @ $2.2B | Pure indirect intake; no BOM or parts |
| Levelpath | $55M Series B | Indirect/enterprise; not hardware BOM |
| Omnea | $138M total | Indirect intake; no direct-materials depth |
| Tonkean | ~$84M → acq. Coupa 2026 | Process orchestration layer, not sourcing-for-parts |
| Globality | ~$356M total | Services scoping, not parts |
Wedge 1 · BOM → Sourcing Timeline wedges/01-bom-timeline/
Vraj's hero insight made real: "the clock doesn't start until someone does it." Right-to-left back-schedule from build date → shows which parts are already LATE.
Most-urgent parts
| Part | Category | Lead | Kickoff by | Send RFQ by | Status |
|---|---|---|---|---|---|
BMI088 | sensor_ic | 18 wk | 2026-06-09 | 2026-06-16 | LATE |
IIS2MDCTR | sensor_ic | 18 wk | 2026-06-09 | 2026-06-16 | LATE |
MAX-8Q | sensor_ic | 18 wk | 2026-06-09 | 2026-06-16 | LATE |
MPRLS0025PA00001A | sensor_ic | 18 wk | 2026-06-09 | 2026-06-16 | LATE |
W25N01GVZEIG TR | sensor_ic | 18 wk | 2026-06-09 | 2026-06-16 | LATE |
STM32H753VIT6 | mcu | 16 wk | 2026-06-23 | 2026-06-30 | LATE |
STM32F405RGT6 | mcu | 16 wk | 2026-06-23 | 2026-06-30 | LATE |
MIC26903YJL-TR | regulator_ic | 14 wk | 2026-07-07 | 2026-07-14 | ACT NOW |
INA219AIDCNR | regulator_ic | 14 wk | 2026-07-07 | 2026-07-14 | ACT NOW |
PCA9685PW,118 | regulator_ic | 14 wk | 2026-07-07 | 2026-07-14 | ACT NOW |
53048-0210 | connector | 10 wk | 2026-08-04 | 2026-08-11 | SOON |
ABM8G-25.000MHZ | crystal | 10 wk | 2026-08-04 | 2026-08-11 | SOON |
BOM: HADES open-source UAV flight controller by Philip Salmony (BSD-3-Clause) · 67 parts · Dual STM32, Bosch BMI088, u-blox MAX-8Q, Honeywell MPRLS
Wedge 2 · RFQ → Bid Comparison wedges/02-rfq-bid-comparison/
Parsed 3 messy quotes (USD email / EUR pipe-delimited / non-itemized lump sum) → normalized, split-award, auto-drafted follow-ups. This is the daily-use revenue wedge.
Itemized comparison (USD normalized)
| Part | Qty | A-Acme (USD) | Lead | B-Globex (USD) | Lead | Award |
|---|---|---|---|---|---|---|
STM32F405RGT6 | 500 | $6.85 | 8 wk | $6.70 | 9 wk | B-Globex ↓$0.15 |
LSM6DSOXTR | 500 | $3.10 | 10 wk | $3.19 | 14 wk | A-Acme ↓$0.09 |
DF13-10P-1.25H | 500 | $0.62 | 6 wk | $0.76 | 8 wk | A-Acme ↓$0.14 |
GRM188R71C104KA01 | 2000 | $0.01 | 4 wk | $0.01 | 5 wk | B-Globex (price tie) |
NCR18650B | 500 | $4.20 | 12 wk | — not quoted | A-Acme (only source) |
Split award vs. sole-source
| Supplier | Lines | Complete? | Total (USD) | Max lead |
|---|---|---|---|---|
| A-Acme sole-source | 5/5 | Yes | $7,407.00 | 12 wk |
| B-Globex sole-source | 4/5 | NO | $5,338.44 | 14 wk |
| Split award | 5/5 | Yes | $7,327.44 | 12 wk crit |
Savings: $79.56 (1.1%) vs. sole-sourcing cheapest complete supplier.
Flags (what blocks a decision)
- MOQ overage: B-Globex quotes DF13-10P-1.25H at MOQ 1000 but RFQ qty is 500 → forced buy of 500 extra
- MOQ overage: A-Acme quotes GRM188R71C104KA01 at MOQ 4000 but RFQ qty is 2000 → forced buy of 2000 extra
- NCR18650B (12 wk) is SINGLE-SOURCED → schedule + allocation risk; qualify a 2nd source
- C-Initech returned a non-itemized lump sum ($7,950 for "complete kit") → auto follow-up drafted requesting line-item breakdown
2 follow-up emails auto-drafted: followup-B-globex-missing-lines.txt, followup-C-initech-itemize.txt
Wedge 3 · Part & Supplier Intelligence wedges/03-part-supplier-intelligence/
Live USITC HTS tariff data (keyless REST API) joined to BOM. Connectors go from 2.7% → 27.7% in a China-sourcing scenario. 20 lines on EOL watch.
Highest tariff-exposed (China scenario)
| Category | HTS Code | MFN Base Rate | China scenario | Delta |
|---|---|---|---|---|
| connector | 8536.69.80.00 | 2.7% | 27.7% | +25pts |
| lithium_cell | 8507.60 | 3.4% | 28.4% | +25pts |
| dc_motor | 8501.31.20.00 | 2.8% | 27.8% | +25pts |
| ceramic_cap | 8532.24.00 | 0.0% | 25.0% | +25pts |
| mcu | 8542.31 | Free | 25.0% | +25pts |
| resistor | 8533.21 | Free | 25.0% | +25pts |
| pcb_bare | 8534.00 | Free | 25.0% | +25pts |
Base duty = real USITC MFN rates (live API, no key). China scenario = illustrative +25pt Section-301-style adder. Confirm vs. HTS Ch.99 + current EOs before use.
Long-lead criticals (>12 wk)
| Part | Category | Typical lead | EOL watch |
|---|---|---|---|
BMI088 | sensor_ic | ~18 wk | Yes |
IIS2MDCTR | sensor_ic | ~18 wk | Yes |
MAX-8Q | sensor_ic | ~18 wk | Yes |
MPRLS0025PA00001A | sensor_ic | ~18 wk | Yes |
STM32H753VIT6 | mcu | ~16 wk | Yes |
STM32F405RGT6 | mcu | ~16 wk | Yes |
MIC26903YJL-TR | regulator_ic | ~14 wk | Yes |
20 of 67 lines in EOL-prone categories. Production wires a live PCN/EOL feed (SiliconExpert / Z2Data) here. 7 parts need manual category classification.
Open Questions from docs/08-open-questions.md
A. Competitive (highest priority)
- Does Vraj know about Lumari (YC X25)? The genuinely-agentic, direct-materials, head-on competitor with an ex-Google/Tesla/Amazon/Stripe team — he didn't mention it. How do we beat it?
- Is the unnamed "$1.5M Lab126" competitor LightSource? What exactly did Zoox dislike about it — durable wedge or temporary feature gap?
- What's the honest moat in 18 months when LightSource/Cofactr ship agents?
B. Market & Data
- Verify the ICP count — how many hardware companies fit ~$30M+ direct spend? (weakest number in the thesis)
- Data cold-start plan: which paid feeds (SiliconExpert / Z2Data / Nexar) at what cost?
- Is the "$1.5M / 70%-manual" figure corroborated or single-source from Vraj's memory?
C. Product
- Which wedge does the drone partner actually want first?
- Does Vraj agree on email-native, no supplier portal? (research says this is existential)
- What's in his current prototype — and what's the architecture? (take the ~30-min demo)
D. Go-to-Market
- Status of the drone design-partner contract? Signed = real; verbal = pipeline.
- How deep is the warm pipeline beyond partner #1 — names, not vibes?
- Align on paid pilots with success criteria, not "6 months free."
E. Structure (never discussed — must be)
- Equity split and vesting (standard 4-yr / 1-yr cliff?)
- Roles / decision rights — CEO/CPO (Vraj) + CTO (Ian) was stated; what are the actual boundaries?
- IP assignment — the prototype is Vraj's; how is prior + future IP handled?
- Salary trigger — when does income start? Vraj is living on savings.
F–G. Fit & Portfolio
- After the co-working day: is the working relationship right? That's the real point of the day.
- Is this the bet Ian wants vs. other ICG Labs options + a possible salaried return? (Vraj offered a Zoox referral.) A co-founder commitment is 8–10 years.
Roadmap from docs/07-roadmap.md
Phase 0 — Co-working day (now) ✓
Lead with Wedge 1 (BOM→timeline). Show the through-line: one real BOM → three agents. Goal = test compatibility + answer Vraj's agentic-depth reservation.
Phase 1 — One wedge, one partner (0–3 mo)
Harden Wedge 2 (RFQ→bid) for the drone partner. LLM doc extractor + Gmail/Outlook ingest. BYOK pricing API. Charge for the pilot.
Phase 2 — Stitch + expand (3–9 mo)
BOM → timeline → auto-RFQ → bid → award as one flow. Add agentic sends. 2–3 more partners. First paid annual contract ($50–100K).
Phase 3 — System of record + raise (9–18 mo)
BOM / quote history / supplier graph accruing = real switching cost. Add clear-to-build tracker. SAFE / pre-seed on paying logos.
Phase 4 — AI-native ERP expansion (18+ mo)
Expand from sourcing into adjacent modules (inventory, supply/demand planning). "Be the SAP these companies never had to buy." This is where the market becomes venture-scale. The beachhead alone is a Series A/B, not a unicorn — the thesis depends on this expansion.
The four disciplines that make this work
- N=1 useful, always. Every phase must help the first customer with zero network.
- Email-native, no portal. Non-negotiable — it's what killed Scoutbee and sank Ariba adoption.
- Charge early. Paid pilots with success criteria, not free-forever.
- Speed. The competitive window is ~12–24 months.