What I'm building
Live products you can use today, the open-source tools underneath them, and the enterprise systems I ship as Tech Lead.
What I build across
Live products
I don't just architect for others. I put my own products in people's hands.
Two of these are becoming something bigger, where most of my nights-and-weekends energy goes right now.
Provael
Red-team & assurance for physical AI
Attacks vision-language-action (VLA) robot policies in simulation and reports an attack-success rate with a 95% confidence interval and a benign control. I authored the Embodied AI Security Top 10. The embodied twin of my agent-security work.
MannSetu
Voice-first AI mental wellness for India
Hindi / English / Hinglish voice companion with real-time voice-tone emotion analysis and CBT-based guidance. Built to close the therapy access gap for 18-35-year-olds: zero wait-time, private, DPDP-compliant, data on Indian servers.
Also live
Shipped and in people's hands: a solo consumer tool, and the enterprise platform I architect as Tech Lead.
Why Can't We Have An Agent For This?
Consumer AI feasibility analyzer
Type any everyday problem, get an AI-agent feasibility verdict with real competitor research and kill predictions. Solo-built, launched during Holi 2026 with zero marketing spend.
Agentify
Enterprise multi-agent platform
15+ specialized agents for full SDLC automation, PRD to deployment. Cascading-router cost strategy, E2B micro-VM sandboxing per agent, multi-tenant SaaS with Stripe billing. Architected and led end-to-end.
Open source
The tools underneath the products
Agent-security scanners, firewalls, and memory I ship and dogfood on my own agents, plus the evaluation plugin that rates them.
As of 2026-07-25 · Metrics auto-synced from GitHub.
agent-airlock
Runtime firewall for AI agents. Ghost-argument stripping, strict type validation, PII masking, RBAC, E2B sandboxing, network airgap, circuit breaker, cost tracking.
verdict
Universal quality judge for Claude Code. 7-dimension scoring (correctness, completeness, adherence, efficiency, safety), configurable rubrics, threshold blocking, auto-hooks.
Enterprise engineering
The production track record behind the products
Tech Lead at Attri.ai. The flagship internal platform, plus six anonymized customer engagements across six verticals. Every number is sourced from internal evidence; client names withheld under confidentiality.
Built the Org-Wide CI/CD Floor - 200-Repo Audit → Shipped Kit
Platform Tech Lead - design, build, operate
Ran an org-wide CI/CD audit across all active repositories at Attri.ai. Initial findings: 1,135 CRITICAL across 10 audit dimensions - 81% of repos had no branch protection, 72 repos had verified live secret leaks (332 OpenAI keys + 88 AWS access tokens + 30 GitHub PATs in git history), 0% had GitHub secret-scanning or Dependabot security updates enabled, 48% of merged PRs had zero review records. Then designed and shipped `attri-dev-kit` - semver-versioned (current v1.7.0), self-testing, one-line YAML opt-in, language-aware across Python + TypeScript + Rust + Terraform + Shell + C#.
Outcomes
- Org-wide adoption rolling out across 208 repos via a single repository-rollout tracker
- AI-specific guardrails: hallucinated-import detection, swallowed-exception flags across 5 languages, unjustified-lint-disable hard-blocks, AI-author signature failure escalation, test-delta gate on production code
- Self-gating: the kit runs against itself on every PR - we don't ship a version that fails its own check
- Framing: 'AI as a risk-multiplier' - pre-AI the cost of a careless commit was bounded by typing speed; post-AI a frontier model produces 200 lines of plausible-but-wrong code in 30 seconds. The kit absorbs the multiplier so individual engineers don't have to remember to defend against it
Six anonymized engagements across six verticals
Industry vertical, role, scope, and outcome across 17 months at Attri.ai. Client names withheld under contractual confidentiality. Every number, scope item, and outcome below is sourced from internal evidence (commit history, PR review counts, authored docs, calendar collaborator graph).
50-State Production API Integration + MS Graph OAuth2 Migration
Tech Lead - end-to-end ownership
Owned the integration of a regulated 50-state premium / tax / coverage-type API across all US jurisdictions for a commercial insurance carrier. Drove the cutover from legacy SMTP to Microsoft Graph + OAuth2 client-credentials for compliance-sensitive transactional notifications. Diagnosed and resolved a vendor auth ambiguity (Secret-ID vs Secret-Value confusion, AADSTS7000215) in one business day - validated token endpoint (200 OK) + Graph sendMail (202 Accepted) production-ready before EOD.
Production Ops Portal v4 - Jobs / Timesheets / ERP Integrations
Senior IC + customer principal - 17 months
Owned the production operations portal for a US general contractor: jobs management, labor timesheets, hours summary export, audit log, trucking, supplier + products DB, vendor portal, daily recap emails with material/equipment cost columns. Built integrations with industry-standard inventory + construction-management platforms (end-to-end OAuth flows, sandboxed test envs, prod cutover). Drove a vendor-diversity automation pipeline: matching company records against a public diversity-program registry with confidence thresholds + manual override UI.
Greenfield Healthcare AI Platform - Both Halves in 6 Weeks
Tech Lead - end-to-end greenfield
End-to-end ownership of a new healthcare AI product launch - both halves of the stack (FastAPI backend + React frontend). Brought to production from empty repo to v1 cutover in 6 weeks. Authored the Infrastructure & Compliance Audit identifying 8 critical pre-launch gaps: unredacted PHI passed to model providers, subscription ownership mismatch, missing BAA/SLA/IP clauses in vendor contracts, plaintext secrets in App Service settings, Postgres `log_statement=all` logging PHI parameters, missing diagnostic settings.
121-Issue Codebase Audit in 30 Days + Remediation Roadmap
Senior auditor - code review, architecture, remediation lead
Inherited a struggling MERN-stack platform serving K-12 students. Ran a parallel-AI-agent code review across all three repos (frontend, backend, AI service) producing 121 categorised issues, 23 of them Critical including: auth bypass, concurrency-driven data corruption, payment gaps, zero automated test coverage. Authored a 16-section Master Engineering Plan: Mongo→Postgres migration, AI question-generation rebuild, Bayesian mastery / IRT engine fixes, infra modernization, observability from zero, COPPA/FERPA/PCI DSS compliance roadmap, mobile (React Native / Expo) strategy.
Enterprise Claude AI Audit + Observability Platform
Platform Tech Lead - system design, security, runbooks
Designed and operates a Claude-based audit + observability platform running inside the firm's own Azure tenant - captures every prompt, response, and tool use into private Azure PostgreSQL in the US, exposed through a private API for compliance reviewers. Designed for ABA Model Rule 1.6 confidentiality + ABA Formal Opinion 512 (generative-AI ethics). Three-party model: model vendor (Anthropic) + cloud platform vendor (Attri) + IT partner (managed-services provider).
Drone-Inspection POC - Vendor Selection + KMZ Verification
Technical Lead - POC scope, vendor eval, ingestion design
Defined the POC scope for an automated drone-inspection pipeline for a US commercial real-estate buyer. Required KML/KMZ exportable flight-mission file + Smart Oblique capture for repeatability. Personally verified `.kmz` flight-path data (`waylines.wpml`) - 'the golden ticket' for reproducible captures. Confirmed multi-format deliverables (OBJ + LAS/LAZ + DXF). Requested AT (aerial triangulation) / Block-Exchange XML for centimeter-grade accuracy. Authorized capture; designed Phase-2 AI ingestion pipeline.
Why Can't We Have An Agent For This?
My solo-built consumer tool. Type a problem below to get the full agent-feasibility analysis on the live product in a new tab.
Why can't we have an agent for this?
Type any problem you wish an AI agent would do for you. Get a brutally honest roast, viability score, competitive landscape, open-source alternatives, agent-readiness scorecard, and a CLAUDE.md scaffold you can drop into Cursor - in ~60 seconds.
Or try one of these
A full agent-feasibility report
Every roast returns the same eight-section structure - quick to skim, brutal where it needs to be, deterministic agent-readiness score at the end.
- 1Verdict + 1-10 scoreTier from "Build it yesterday" to "Don't bother"
- 2Viability sub-scoresMarket demand · feasibility · competition · monetization · disruption risk · fun factor
- 3Pros & consWhat's going for it · what's against it · what kills it
- 4Who you're up againstReal competitors with positioning + threat level
- 5Open-source alternativesWhat you could fork instead of building from zero
- 6Big-AI killer timelineWho absorbs your idea + when + your survival strategy
- 7Build estimateSolo-dev time · team size · cost · suggested tech stack
- 8CLAUDE.md scaffoldDrop-in starter spec for Claude Code / Cursor
The Production Agent
Newsletter
Weekly lessons from running 15+ AI agents in production. Governance, security, memory, cost optimization. No demos. Systems that work.
- ✓ Agent orchestration & governance patterns
- ✓ LLM cost optimization strategies
- ✓ Security & memory architecture deep-dives
- ✓ Real production war stories & lessons
Free forever. No spam. Unsubscribe anytime.
Building something in AI?
Whether you're hiring a builder who ships, want to compare notes on physical-AI security or agent governance, or need a production AI platform architected, let's talk.