Sai Arun Gunuganti — Lead Software Engineer. I design and build agentic products that do real work in sectors where mistakes cost money, care, or trust.
Every product here runs on one spine: rules in code for money, eligibility, deadlines and safety — never decided by a model; human approval at every consequential step; and a sealed, verifiable receipt for everything the agent did.
In production, with real users and paying-customer discipline.
Spend authorization for AI agents
The problem: AI agents can spend money, but nothing checks the owner's limits before the spend — or proves afterwards what was approved, by whom, and why.
Sealgraf is the authorization layer: the agent checks owner-set limits before spending, obtains the owner's approval on a surface the agent cannot fake, and seals a verifiable receipt for every transaction. Integration packages for six agent frameworks — Dify, CrewAI, LangChain, n8n, Composio, and the Vercel AI SDK — each live-tested end-to-end with server-verified approval loops.
One product per sector, built from full-sector problem research — the company and all its competitors, not one company's complaint list. Working prototypes: complete journeys, rules enforced in code, approval gates on, synthetic data only.
Workplace mental health · EAP access
EAP utilization runs 3–8% because the access funnel leaks before care: undiscoverable benefits, eligibility traps, black-box matching, silent session caps.
An agentic front door that completes the journey — eligibility verified without hard denials, the benefit explained with citations, steerable matching with shown reasoning, booking behind approval gates, session balance tracked from a ledger, and a member-visible privacy ledger. Crisis signals route to humans (988) before anything else; employers see aggregates only, behind a cohort floor.
Short-term rentals & travel
After booking is where guests get hurt: support failures, hijacked-chat payment scams, host cancellations, and protection programs that are explicitly not insurance.
A guest-side recovery agent: booking verified, evidence sealed in a vault at booking time, payment messages authenticated against hard rules in code (it will never call an off-platform payment "safe"), claims filed inside platform windows behind approval, refunds tracked to recovered.
Real-estate portals & rentals
Search works; everything after fails — wrong data with no correction path, scraped scam listings, screening errors, and fees on invisible rails.
The seeker's verification agent: any listing verified with a receipt — ownership cross-checks, photo provenance, price-vs-market — with verdicts of VERIFIED / UNVERIFIABLE / RED FLAGS and the evidence shown. It never prints "safe." Portable application profile, a fee ledger, and a hard budget gate where any change invalidates the approval.
Local-services marketplaces
Pros pay for leads, not jobs: ghost leads, shared-lead races, and charge triggers outside the pro's control — while consumers have no neutral record when work goes wrong.
A two-sided evidence layer. Pros: every lead charge audited against the platform's own rules as of the charge date, dispute packets assembled, filings made only with the pro's approval — and a filing only counts as filed with a confirmation artifact. Consumers: a signed scope record before work begins.
Consumer banking & shopping rewards
Rewards flip from "pending" to "Ineligible $0.00" with no reason; disputes are decided against documented cardholders; deadlines hide in statement fine print.
A member-side recovery agent: a WHY-ledger where every denial carries an evidenced, rule-cited reason — or is honestly marked reason unknown — and a dispute copilot whose statutory clocks (FCBA, Reg E, FCRA) run in code from the correct anchor dates. No bank-credential vault exists anywhere in its data model.
One platform, five products
State-machine orchestrator with per-state tool allowlists enforced in code · append-only, hash-chained ledger · citation-or-refuse machinery · safety and honesty layers · sandbox-first adapters that degrade honestly where platforms offer no access. 239 tests across the five products, all passing.
Memory engine for AI agents · Open source
A local-first memory engine: SQLite + sqlite-vec storage, a bitemporal graph of events with transitions as first-class data, hybrid retrieval, MCP / REST / CLI surfaces, and forget-by-default TTLs — memory that can explain and expire itself.
Agent assurance · Prototype
An assurance platform for AI agents: a detection engine with financial guardrails, compliance and agent-behavior checks, and an MCP trust-layer proxy that issues allow / warn / block verdicts before an agent's action lands.