01
Billing at the edge
Credits and API keys for long-running hosted agents on Cloudflare Workers. Key issuance, metering, accounting, quota. Correctness under concurrency.
Samarth Gugnani
Software engineer — backend, data, and agent systems
Samarth Gugnani is a software engineer with six years of remote work for distributed teams. He built Eliza billing on Cloudflare Workers — the credits and API-key platform thousands of hosted AI agents ran on — and independently built a GTM data stack from scratch and ran it in production. He is not a blockchain developer.
Billing that holds under concurrency. Pipelines you can open at the source. Agents that abstain when they cannot prove the row.
Remote · six years remote for distributed teams · overlap hours US, EU or APAC
Proof, not adjectives
- 15,000+
- GitHub stars on elizaOS. Core maintainer. 100+ contributors.
- Thousands
- of hosted agents billed on Cloudflare Workers. Correctness under concurrency.
- 351
- companies live at the ATS. 456 contacts. Source API on every row.
- 60%
- faster data layer after resolving one creator across five platforms.
The old way is a guess
Source at the API, or it does not ship.
Stop
- Cached indexes treated as the list
- A canvas until the vendor does not expose the endpoint
- An LLM that writes a CRM that was never on the page
- Billing that is eventually consistent
Start
- Re-check at the source before use
- Write against the real API
- Drop the claim. Abstain when confidence is low
- Quota that holds when thousands of agents hit the edge at once
What the work actually is
Six systems. All shipped.
02
Disagreeing sources, one schema
TikTok’s private API, then the same creator across YouTube, Twitter/X, Instagram and Spotify. Data-layer response times down 60%.
03
Signal verified at the source
Five ATS APIs. 351 companies re-checked live. Cached indexes go stale. Every row stamped with source API and timestamp.
04
Installs, not vendor names
DNS, headers, cookies, resource hosts. HubSpot portal ID and GA4 measurement ID as proof. A name in page copy is rejected with a reason.
05
Extraction that abstains
13 chained prompts. 5,000+ accounts. Any claim not on the source page is dropped. Extractors abstain when confidence is low.
06
A human on the line
Next.js operator UI. Accept, reject, or correct before anything ships. What an agent may settle alone is decided up front.
How it holds
Three rules. Not a stack of tools.
1
Prove it at the source
ATS, DNS, headers, the page itself. If it is not there, it is not a row.
2
Abstain when unsure
Extractors drop the claim. Unverified contacts are quarantined with a reason.
3
A human signs off
What an agent may settle alone is decided up front. The rest waits.
Selected work
Open the writeup.
Credits and API-key billing at Eliza Labs
The platform thousands of hosted agents ran on. I did not found Eliza. I owned billing.
A GTM data stack, crawler to operator UI
Live ATS verification, first-party tech detection, enrichment that quarantines, LLM extraction that abstains.
Entity resolution across five social APIs
Five sources that agreed on nothing. One schema. Data layer 60% faster.
Experience
Title, company, dates.
- Software engineer, Independent
- Senior Software Engineer, Eliza Labs
- Full-stack engineer, Camp Network
- Full-stack engineer, Shield (a16z-backed)
- CTO, SignAssist
- Founder, Intelli Chains
Earlier, 2020–2021, dated only: Cryption Network, YFDAI Finance, Unbox Innovations, Google Summer of Code (Amahi). B.Tech Computer Science, Chitkara University. I am not a blockchain developer.
If the work is this, write.
Samarth Gugnani builds backend and data infrastructure and the LLM systems on top of them: billing that stays correct under concurrency, pipelines that turn messy third-party APIs into records you can trust, and the guardrails that decide where an agent runs alone and where a human signs off.
Email Samarth Gugnani