Ask HN: Architecting audit-grade ESG platforms – AI assistants vs. human CTOs

Background: I'm a solo technical founder building Velumin, a carbon accounting platform for Fortune 500 compliance (CSRD, BRSR, GHG Protocol).

The challenge: ESG platforms need: - Deterministic calculations (auditors reject "AI math") - Immutable audit trails (SOX/SOC2 requirements) - Multi-jurisdictional compliance (EU CSRD, India BRSR, US SEC) - Real-time anomaly detection + AI document generation

*My experiment:* I used Cursor, GitHub Copilot, and Amazon Q (Kiro) to architect the entire stack, guided by a structured "WAR-MODE" prompt covering: 1. Technical architecture (multi-region, event sourcing, circuit breakers) 2. ESG methodology (GHG Protocol validators, uncertainty quantification) 3. Regulatory engines (BRSR/CSRD/SEC automation) 4. Product/UX (role-based onboarding, supplier agent, no-code workflows)

*AI correctly identified:* "Never use LLMs for emission calculations—auditors will reject it" "Implement WORM storage for audit trails, not 'agent memory'" "Multi-model strategy: GPT-4V for OCR, Claude for reports, rules for compliance" "India-first BRSR compliance = competitive moat"

*What I'm unsure about:* - Are there architectural anti-patterns AI tools systematically miss? - For compliance-critical systems, is AI review a complement or substitute for human CTOs? - What's the right balance of AI-generated architecture vs. human validation?

*For experienced CTOs/architects:* What would you want to validate in a system like this that AI likely couldn't catch? And conversely, are there areas where AI review is now legitimately superior to human review (e.g., exhaustive checklist coverage)?

I'm happy to share: - The full WAR-MODE prompt structure (so you can adapt it) - Our architecture decisions and trade-offs - Specific gaps we're worried about

Curious to hear from folks building audit-grade or compliance-heavy systems.

▲

westurner 9 hours ago | parent [-]

Some forms of carbon are worse than others but carbon mass doesn't account for the difference in impact. Aren't there additional externalities to account for in addition to just carbon?

On whether ESG is worth the time (compared to blindly investing in a universe of stocks that look good on paper relative to other assets only because they're dumping external costs onto everyone without accountability):

"Companies with good ESG scores pollute as much as low-rated rivals" (2023) https://news.ycombinator.com/item?id=36980661

How should carbon accounting account for a process that generates porous graphene filters that capture CO2 carbon out of CO2?

	▲	Jayeshkumbhar 8 hours ago \| parent [-]
		OP here — really appreciate these questions because they get at the real limitations of carbon accounting frameworks. 1. "Carbon ≠ carbon": different gases, different externalities Totally agree. CO₂ mass alone is a simplification. That's why GHG Protocol uses GWP factors to convert different gases into CO₂e: - CH₄: 28–34× CO₂ - N₂O: 265–298× - SF₆/HFCs: 10,000×+ But even GWP misses important dimensions: - Timing effects (short-lived vs. long-lived gases) - Toxicity and pollution - Ozone impacts - Ecosystem and social externalities So in our system, carbon accounting is just the starting layer. CSRD already forces companies to track water, biodiversity, pollution, and circularity on top of climate (ESRS E2-E5). 2. Re: ESG ratings not correlating with lower emissions Fully agree with the critique. Most ESG scores measure: - Disclosures instead of actual performance - Policies instead of physics - Governance/social weighting that dilutes environmental signals That's why we avoid "ESG scores" completely. We follow: - Strict GHG Protocol methods - Audit-grade emission-factor calculations - CSRD/BRSR/SEC climate-rule compliance The 2023 study you cited is exactly why deterministic calculation matters more than ratings. 3. On porous graphene and carbon-capture edge cases This is where things get interesting. Under GHG Protocol: - Manufacturing the filter → positive emissions (Scope 1/2/3) - Capturing CO₂ → potential removal - But: only counts as removal if storage is permanent (>100 yrs) and third-party verified (e.g., Puro.earth, CDR.fyi) - Temporary use (e.g., carbonation) is not removal—just delayed re-emission In our accounting model we separate: - Emissions (tCO₂e released) - Avoidance (vs. baseline) - Removals (atmospheric drawdown) - Permanence categories (geological, mineralization, engineered, biomass) - Uncertainty ranges (required under CSRD ESRS E1) Your graphene example is exactly the type of nuance that standard ESG dashboards usually ignore. 4. Genuine curiosity Do you work in carbon accounting, lifecycle analysis, or climate methodology? Your questions suggest real hands-on experience with the edge cases. We're building Velumin's methodology to handle exactly these scenarios—would love to hear more about your experience if you're open to it. --- Side note: Still interested in the original topic — for compliance-heavy systems, I'm trying to understand where experienced engineers think AI architecture review breaks down vs. where it actually outperforms humans (especially in checklist coverage).